chore: delete unused files and outdated GUI_DEMO documentation

feat(gui): display face ROIs in comparison results

refactor(bridge): pad detected face bounding boxes uniformly

.gitattributes

@@ -0,0 +1,4 @@
+models/facenet.onnx filter=lfs diff=lfs merge=lfs -text
+models/retinaface.mnn filter=lfs diff=lfs merge=lfs -text
+models/retinaface.onnx filter=lfs diff=lfs merge=lfs -text
+models/facenet.mnn filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -2,3 +2,6 @@
 /target
 .direnv
 *.jpg
+face_net.onnx
+.DS_Store
+*.cache

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "rfcs"]
+	path = rfcs
+	url = git@github.com:aftershootco/rfcs.git

Cargo.toml

@@ -1,14 +1,27 @@
 [workspace]
-members = ["ndarray-image", "ndarray-resize", ".", "bounding-box"]
+members = [
+    "ndarray-image",
+    "ndarray-resize",
+    ".",
+    "bounding-box",
+    "ndarray-safetensors",
+    "sqlite3-ndarray-math",
+    "ndcv-bridge",
+    "bbox",
+]
 
 [workspace.package]
 version = "0.1.0"
 edition = "2024"
 
-[patch."https://github.com/uttarayan21/mnn-rs"]
-mnn = { path = "/Users/fs0c131y/Projects/aftershoot/mnn-rs" }
+[patch.crates-io]
+linfa = { git = "https://github.com/relf/linfa", branch = "upgrade-ndarray-0.16" }
+linfa-clustering = { git = "https://github.com/relf/linfa", branch = "upgrade-ndarray-0.16" }
 
 [workspace.dependencies]
+divan = { version = "0.1.21" }
+ndarray-npy = "0.9.1"
+serde = { version = "1.0", features = ["derive"] }
 ndarray-image = { path = "ndarray-image" }
 ndarray-resize = { path = "ndarray-resize" }
 mnn = { git = "https://github.com/uttarayan21/mnn-rs", version = "0.2.0", features = [
@@ -22,6 +35,16 @@ mnn-bridge = { git = "https://github.com/uttarayan21/mnn-rs", version = "0.1.0",
 mnn-sync = { git = "https://github.com/uttarayan21/mnn-rs", version = "0.1.0", features = [
     "tracing",
 ], branch = "restructure-tensor-type" }
+nalgebra = { version = "0.34.0", default-features = false, features = ["std"] }
+opencv = { version = "0.95.1" }
+bounding-box = { path = "bounding-box" }
+bytemuck = "1.23.2"
+error-stack = "0.5.0"
+thiserror = "2.0"
+fast_image_resize = "5.2.0"
+img-parts = "0.4.0"
+ndarray = { version = "0.16.1", features = ["rayon"] }
+num = "0.4"
 
 [package]
 name = "detector"
@@ -35,12 +58,11 @@ clap_complete = "4.5"
 error-stack = "0.5"
 fast_image_resize = "5.2.0"
 image = "0.25.6"
-linfa = "0.7.1"
-nalgebra = "0.33.2"
+nalgebra = { workspace = true }
 ndarray = "0.16.1"
 ndarray-image = { workspace = true }
 ndarray-resize = { workspace = true }
-rusqlite = { version = "0.37.0", features = ["modern-full"] }
+rusqlite = { version = "0.37.0", features = ["functions", "modern-full"] }
 tap = "1.0.1"
 thiserror = "2.0"
 tokio = "1.43.1"
@@ -53,6 +75,39 @@ bounding-box = { version = "0.1.0", path = "bounding-box" }
 color = "0.3.1"
 itertools = "0.14.0"
 ordered-float = "5.0.0"
+ort = { version = "2.0.0-rc.10", default-features = false, features = [
+    "std",
+    "tracing",
+    "ndarray",
+    "cuda",
+] }
+ndarray-math = { git = "https://git.darksailor.dev/servius/ndarray-math", version = "0.1.0" }
+ndarray-safetensors = { version = "0.1.0", path = "ndarray-safetensors" }
+sqlite3-ndarray-math = { version = "0.1.0", path = "sqlite3-ndarray-math" }
+
+# GUI dependencies
+iced = { version = "0.13", features = ["tokio", "image"] }
+rfd = "0.15"
+futures = "0.3"
+imageproc = "0.25"
+linfa = "0.7.1"
+linfa-clustering = "0.7.1"
 
 [profile.release]
 debug = true
+
+[features]
+ort-cuda = []
+ort-coreml = ["ort/coreml"]
+ort-tensorrt = ["ort/tensorrt"]
+ort-tvm = ["ort/tvm"]
+ort-openvino = ["ort/openvino"]
+ort-directml = ["ort/directml"]
+mnn-metal = ["mnn/metal"]
+mnn-coreml = ["mnn/coreml"]
+
+default = ["ort-cuda"]
+
+[[test]]
+name = "test_bbox_replacement"
+path = "test_bbox_replacement.rs"

Makefile.toml

@@ -0,0 +1,38 @@
+[tasks.convert]
+dependencies = ["convert_facenet", "convert_retinaface"]
+workspace = false
+
+[tasks.convert_facenet]
+command = "MNNConvert"
+args = [
+    "-f",
+    "ONNX",
+    "--modelFile",
+    "models/facenet.onnx",
+    "--MNNModel",
+    "models/facenet.mnn",
+    "--fp16",
+    "--bizCode",
+    "MNN",
+]
+workspace = false
+
+[tasks.convert_retinaface]
+command = "MNNConvert"
+args = [
+    "-f",
+    "ONNX",
+    "--modelFile",
+    "models/retinaface.onnx",
+    "--MNNModel",
+    "models/retinaface.mnn",
+    "--fp16",
+    "--bizCode",
+    "MNN",
+]
+workspace = false
+
+[tasks.gui]
+command = "cargo"
+args = ["run", "--release", "--bin", "gui"]
+workspace = false

README.md

@@ -1,3 +1,227 @@
-# Face Detection
+# Face Detection and Embedding
 
-Rust programs to do face detection and face embedding
+A high-performance Rust implementation for face detection and face embedding generation using neural networks.
+
+## Overview
+
+This project provides a complete face detection and recognition pipeline with the following capabilities:
+
+- **Face Detection**: Detect faces in images using RetinaFace model
+- **Face Embedding**: Generate face embeddings using FaceNet model
+- **Multiple Backends**: Support for both MNN and ONNX runtime execution
+- **Hardware Acceleration**: Metal, CoreML, and OpenCL support on compatible platforms
+- **Modular Design**: Workspace architecture with reusable components
+
+## Features
+
+- 🔍 **Accurate Face Detection** - Uses RetinaFace model for robust face detection
+- 🧠 **Face Embeddings** - Generate 512-dimensional face embeddings with FaceNet
+- ⚡ **High Performance** - Optimized with hardware acceleration (Metal, CoreML)
+- 🔧 **Flexible Configuration** - Adjustable detection thresholds and NMS parameters
+- 📦 **Modular Architecture** - Reusable components for image processing and bounding boxes
+- 🖼️ **Visual Output** - Draw bounding boxes on detected faces
+
+## Architecture
+
+The project is organized as a Rust workspace with the following components:
+
+- **`detector`** - Main face detection and embedding application
+- **`bounding-box`** - Geometric operations and drawing utilities for bounding boxes
+- **`ndarray-image`** - Conversion utilities between ndarray and image formats
+- **`ndarray-resize`** - Fast image resizing operations on ndarray data
+
+## Models
+
+The project includes pre-trained neural network models:
+
+- **RetinaFace** - Face detection model (`.mnn` and `.onnx` formats)
+- **FaceNet** - Face embedding model (`.mnn` and `.onnx` formats)
+
+## Usage
+
+### Basic Face Detection
+
+```bash
+# Detect faces using MNN backend (default)
+cargo run --release detect path/to/image.jpg
+
+# Detect faces using ONNX Runtime backend
+cargo run --release detect --executor onnx path/to/image.jpg
+
+# Save output with bounding boxes drawn
+cargo run --release detect --output detected.jpg path/to/image.jpg
+
+# Adjust detection sensitivity
+cargo run --release detect --threshold 0.9 --nms-threshold 0.4 path/to/image.jpg
+```
+
+### Face Comparison
+
+Compare faces between two images by computing and comparing their embeddings:
+
+```bash
+# Compare faces in two images
+cargo run --release compare image1.jpg image2.jpg
+
+# Compare with custom thresholds
+cargo run --release compare --threshold 0.9 --nms-threshold 0.4 image1.jpg image2.jpg
+
+# Use ONNX Runtime backend for comparison
+cargo run --release compare -p cpu image1.jpg image2.jpg
+
+# Use MNN with Metal acceleration
+cargo run --release compare -f metal image1.jpg image2.jpg
+```
+
+The compare command will:
+1. Detect all faces in both images
+2. Generate embeddings for each detected face
+3. Compute cosine similarity between all face pairs
+4. Display similarity scores and the best match
+5. Provide interpretation of the similarity scores:
+   - **> 0.8**: Very likely the same person
+   - **0.6-0.8**: Possibly the same person
+   - **0.4-0.6**: Unlikely to be the same person
+   - **< 0.4**: Very unlikely to be the same person
+
+### Backend Selection
+
+The project supports two inference backends:
+
+- **MNN Backend** (default): High-performance inference framework with Metal/CoreML support
+- **ONNX Runtime Backend**: Cross-platform ML inference with broad hardware support
+
+```bash
+# Use MNN backend with Metal acceleration (macOS)
+cargo run --release detect --executor mnn --forward-type metal path/to/image.jpg
+
+# Use ONNX Runtime backend
+cargo run --release detect --executor onnx path/to/image.jpg
+```
+
+### Command Line Options
+
+```bash
+# Face detection with custom parameters
+cargo run --release detect [OPTIONS] <IMAGE>
+
+Options:
+  -m, --model <MODEL>              Custom model path
+  -M, --model-type <MODEL_TYPE>    Model type [default: retina-face]
+  -o, --output <OUTPUT>            Output image path
+  -e, --executor <EXECUTOR>        Inference backend [mnn, onnx]
+  -f, --forward-type <FORWARD_TYPE> MNN execution backend [default: cpu]
+  -t, --threshold <THRESHOLD>      Detection threshold [default: 0.8]
+  -n, --nms-threshold <NMS_THRESHOLD> NMS threshold [default: 0.3]
+```
+
+### Quick Start
+
+```bash
+# Build the project
+cargo build --release
+
+# Run face detection on sample image
+just run
+# or
+cargo run --release detect ./1000066593.jpg
+```
+
+## Hardware Acceleration
+
+### MNN Backend
+
+The MNN backend supports various execution backends:
+
+- **CPU** - Default, works on all platforms
+- **Metal** - macOS GPU acceleration
+- **CoreML** - macOS/iOS neural engine acceleration
+- **OpenCL** - Cross-platform GPU acceleration
+
+```bash
+# Use Metal acceleration on macOS
+cargo run --release detect --executor mnn --forward-type metal path/to/image.jpg
+
+# Use CoreML on macOS/iOS
+cargo run --release detect --executor mnn --forward-type coreml path/to/image.jpg
+```
+
+### ONNX Runtime Backend
+
+The ONNX Runtime backend automatically selects the best available execution provider based on your system configuration.
+
+## Development
+
+### Prerequisites
+
+- Rust 2024 edition
+- MNN runtime (automatically linked)
+- ONNX runtime (for ONNX backend)
+
+### Building
+
+```bash
+# Standard build
+cargo build
+
+# Release build with optimizations
+cargo build --release
+
+# Run tests
+cargo test
+```
+
+### Project Structure
+
+```
+├── src/
+│   ├── facedet/             # Face detection modules
+│   │   ├── mnn/            # MNN backend implementations
+│   │   ├── ort/            # ONNX Runtime backend implementations
+│   │   └── postprocess.rs  # Shared postprocessing logic
+│   ├── faceembed/          # Face embedding modules
+│   │   ├── mnn/            # MNN backend implementations
+│   │   └── ort/            # ONNX Runtime backend implementations
+│   ├── cli.rs              # Command line interface
+│   └── main.rs             # Application entry point
+├── models/                 # Neural network models (.mnn and .onnx)
+├── bounding-box/           # Bounding box utilities
+├── ndarray-image/          # Image conversion utilities
+└── ndarray-resize/         # Image resizing utilities
+```
+
+### Backend Architecture
+
+The codebase is organized to support multiple inference backends:
+
+- **Common interfaces**: `FaceDetector` and `FaceEmbedder` traits provide unified APIs
+- **Shared postprocessing**: Common logic for anchor generation, NMS, and coordinate decoding
+- **Backend-specific implementations**: Separate modules for MNN and ONNX Runtime
+- **Modular design**: Easy to add new backends by implementing the common traits
+
+## License
+
+MIT License
+
+## Dependencies
+
+Key dependencies include:
+
+- **MNN** - High-performance neural network inference framework (MNN backend)
+- **ONNX Runtime** - Cross-platform ML inference (ORT backend)
+- **ndarray** - N-dimensional array processing
+- **image** - Image processing and I/O
+- **clap** - Command line argument parsing
+- **bounding-box** - Geometric operations for face detection
+- **error-stack** - Structured error handling
+
+### Backend Status
+
+- ✅ **MNN Backend**: Fully implemented with hardware acceleration support
+- 🚧 **ONNX Runtime Backend**: Framework implemented, inference logic to be completed
+
+*Note: The ORT backend currently provides the framework but requires completion of the inference implementation.*
+
+---
+
+*Built with Rust for maximum performance and safety in computer vision applications.*

assets/headshots

@@ -0,0 +1 @@
+/Users/fs0c131y/Pictures/test_cases/compressed/HeadshotJpeg

bbox/Cargo.toml

@@ -0,0 +1,13 @@
+[package]
+name = "bbox"
+version.workspace = true
+edition.workspace = true
+
+[dependencies]
+ndarray.workspace = true
+num = "0.4.3"
+serde = { workspace = true, features = ["derive"], optional = true }
+
+[features]
+serde = ["dep:serde"]
+default = ["serde"]

bbox/src/lib.rs

@@ -0,0 +1,708 @@
+pub mod traits;
+
+/// A bounding box of co-ordinates whose origin is at the top-left corner.
+#[derive(
+    Debug, Copy, Clone, PartialEq, PartialOrd, Eq, Hash, serde::Serialize, serde::Deserialize,
+)]
+#[non_exhaustive]
+pub struct BBox<T = f32> {
+    pub x: T,
+    pub y: T,
+    pub width: T,
+    pub height: T,
+}
+
+impl<T> From<[T; 4]> for BBox<T> {
+    fn from([x, y, width, height]: [T; 4]) -> Self {
+        Self {
+            x,
+            y,
+            width,
+            height,
+        }
+    }
+}
+
+impl<T: Copy> BBox<T> {
+    pub fn new(x: T, y: T, width: T, height: T) -> Self {
+        Self {
+            x,
+            y,
+            width,
+            height,
+        }
+    }
+
+    /// Casts the internal values to another type using [as] keyword
+    pub fn cast<T2>(self) -> BBox<T2>
+    where
+        T: num::cast::AsPrimitive<T2>,
+        T2: Copy + 'static,
+    {
+        BBox {
+            x: self.x.as_(),
+            y: self.y.as_(),
+            width: self.width.as_(),
+            height: self.height.as_(),
+        }
+    }
+
+    /// Clamps all the internal values to the given min and max.
+    pub fn clamp(&self, min: T, max: T) -> Self
+    where
+        T: std::cmp::PartialOrd,
+    {
+        Self {
+            x: num::clamp(self.x, min, max),
+            y: num::clamp(self.y, min, max),
+            width: num::clamp(self.width, min, max),
+            height: num::clamp(self.height, min, max),
+        }
+    }
+
+    pub fn clamp_box(&self, bbox: BBox<T>) -> Self
+    where
+        T: std::cmp::PartialOrd,
+        T: num::Zero,
+        T: core::ops::Add<Output = T>,
+        T: core::ops::Sub<Output = T>,
+    {
+        let x1 = num::clamp(self.x1(), bbox.x1(), bbox.x2());
+        let y1 = num::clamp(self.y1(), bbox.y1(), bbox.y2());
+        let x2 = num::clamp(self.x2(), bbox.x1(), bbox.x2());
+        let y2 = num::clamp(self.y2(), bbox.y1(), bbox.y2());
+        Self::new_xyxy(x1, y1, x2, y2)
+    }
+
+    pub fn normalize(&self, width: T, height: T) -> Self
+    where
+        T: core::ops::Div<Output = T> + Copy,
+    {
+        Self {
+            x: self.x / width,
+            y: self.y / height,
+            width: self.width / width,
+            height: self.height / height,
+        }
+    }
+
+    /// Normalize after casting to float
+    pub fn normalize_f64(&self, width: T, height: T) -> BBox<f64>
+    where
+        T: core::ops::Div<Output = T> + Copy,
+        T: num::cast::AsPrimitive<f64>,
+    {
+        BBox {
+            x: self.x.as_() / width.as_(),
+            y: self.y.as_() / height.as_(),
+            width: self.width.as_() / width.as_(),
+            height: self.height.as_() / height.as_(),
+        }
+    }
+
+    pub fn denormalize(&self, width: T, height: T) -> Self
+    where
+        T: core::ops::Mul<Output = T> + Copy,
+    {
+        Self {
+            x: self.x * width,
+            y: self.y * height,
+            width: self.width * width,
+            height: self.height * height,
+        }
+    }
+
+    pub fn height(&self) -> T {
+        self.height
+    }
+
+    pub fn width(&self) -> T {
+        self.width
+    }
+
+    pub fn padding(&self, padding: T) -> Self
+    where
+        T: core::ops::Add<Output = T> + core::ops::Sub<Output = T> + Copy,
+    {
+        Self {
+            x: self.x - padding,
+            y: self.y - padding,
+            width: self.width + padding + padding,
+            height: self.height + padding + padding,
+        }
+    }
+
+    pub fn padding_height(&self, padding: T) -> Self
+    where
+        T: core::ops::Add<Output = T> + core::ops::Sub<Output = T> + Copy,
+    {
+        Self {
+            x: self.x,
+            y: self.y - padding,
+            width: self.width,
+            height: self.height + padding + padding,
+        }
+    }
+
+    pub fn padding_width(&self, padding: T) -> Self
+    where
+        T: core::ops::Add<Output = T> + core::ops::Sub<Output = T> + Copy,
+    {
+        Self {
+            x: self.x - padding,
+            y: self.y,
+            width: self.width + padding + padding,
+            height: self.height,
+        }
+    }
+
+    // Enlarge / shrink the bounding box by a factor while
+    // keeping the center point and the aspect ratio fixed
+    pub fn scale(&self, factor: T) -> Self
+    where
+        T: core::ops::Mul<Output = T>,
+        T: core::ops::Sub<Output = T>,
+        T: core::ops::Add<Output = T>,
+        T: core::ops::Div<Output = T>,
+        T: num::One + Copy,
+    {
+        let two = num::one::<T>() + num::one::<T>();
+        let width = self.width * factor;
+        let height = self.height * factor;
+        let width_inc = width - self.width;
+        let height_inc = height - self.height;
+        Self {
+            x: self.x - width_inc / two,
+            y: self.y - height_inc / two,
+            width,
+            height,
+        }
+    }
+
+    pub fn scale_x(&self, factor: T) -> Self
+    where
+        T: core::ops::Mul<Output = T>
+            + core::ops::Sub<Output = T>
+            + core::ops::Add<Output = T>
+            + core::ops::Div<Output = T>
+            + num::One
+            + Copy,
+    {
+        let two = num::one::<T>() + num::one::<T>();
+        let width = self.width * factor;
+        let width_inc = width - self.width;
+        Self {
+            x: self.x - width_inc / two,
+            y: self.y,
+            width,
+            height: self.height,
+        }
+    }
+
+    pub fn scale_y(&self, factor: T) -> Self
+    where
+        T: core::ops::Mul<Output = T>
+            + core::ops::Sub<Output = T>
+            + core::ops::Add<Output = T>
+            + core::ops::Div<Output = T>
+            + num::One
+            + Copy,
+    {
+        let two = num::one::<T>() + num::one::<T>();
+        let height = self.height * factor;
+        let height_inc = height - self.height;
+        Self {
+            x: self.x,
+            y: self.y - height_inc / two,
+            width: self.width,
+            height,
+        }
+    }
+
+    pub fn offset(&self, offset: Point<T>) -> Self
+    where
+        T: core::ops::Add<Output = T> + Copy,
+    {
+        Self {
+            x: self.x + offset.x,
+            y: self.y + offset.y,
+            width: self.width,
+            height: self.height,
+        }
+    }
+
+    /// Translate the bounding box by the given offset
+    /// if they are in the same scale
+    pub fn translate(&self, bbox: Self) -> Self
+    where
+        T: core::ops::Add<Output = T> + Copy,
+    {
+        Self {
+            x: self.x + bbox.x,
+            y: self.y + bbox.y,
+            width: self.width,
+            height: self.height,
+        }
+    }
+
+    pub fn with_top_left(&self, top_left: Point<T>) -> Self {
+        Self {
+            x: top_left.x,
+            y: top_left.y,
+            width: self.width,
+            height: self.height,
+        }
+    }
+
+    pub fn center(&self) -> Point<T>
+    where
+        T: core::ops::Add<Output = T> + core::ops::Div<Output = T> + Copy,
+        T: num::One,
+    {
+        let two = T::one() + T::one();
+        Point::new(self.x + self.width / two, self.y + self.height / two)
+    }
+
+    pub fn area(&self) -> T
+    where
+        T: core::ops::Mul<Output = T> + Copy,
+    {
+        self.width * self.height
+    }
+
+    // Corresponds to self.x1() and self.y1()
+    pub fn top_left(&self) -> Point<T> {
+        Point::new(self.x, self.y)
+    }
+
+    pub fn top_right(&self) -> Point<T>
+    where
+        T: core::ops::Add<Output = T> + Copy,
+    {
+        Point::new(self.x + self.width, self.y)
+    }
+
+    pub fn bottom_left(&self) -> Point<T>
+    where
+        T: core::ops::Add<Output = T> + Copy,
+    {
+        Point::new(self.x, self.y + self.height)
+    }
+
+    // Corresponds to self.x2() and self.y2()
+    pub fn bottom_right(&self) -> Point<T>
+    where
+        T: core::ops::Add<Output = T> + Copy,
+    {
+        Point::new(self.x + self.width, self.y + self.height)
+    }
+
+    pub const fn x1(&self) -> T {
+        self.x
+    }
+
+    pub const fn y1(&self) -> T {
+        self.y
+    }
+
+    pub fn x2(&self) -> T
+    where
+        T: core::ops::Add<Output = T> + Copy,
+    {
+        self.x + self.width
+    }
+
+    pub fn y2(&self) -> T
+    where
+        T: core::ops::Add<Output = T> + Copy,
+    {
+        self.y + self.height
+    }
+
+    pub fn overlap(&self, other: &Self) -> T
+    where
+        T: std::cmp::PartialOrd
+            + traits::min::Min
+            + traits::max::Max
+            + num::Zero
+            + core::ops::Add<Output = T>
+            + core::ops::Sub<Output = T>
+            + core::ops::Mul<Output = T>
+            + Copy,
+    {
+        let x1 = self.x.max(other.x);
+        let y1 = self.y.max(other.y);
+        let x2 = (self.x + self.width).min(other.x + other.width);
+        let y2 = (self.y + self.height).min(other.y + other.height);
+        let width = (x2 - x1).max(T::zero());
+        let height = (y2 - y1).max(T::zero());
+        width * height
+    }
+
+    pub fn iou(&self, other: &Self) -> T
+    where
+        T: std::cmp::Ord
+            + num::Zero
+            + traits::min::Min
+            + traits::max::Max
+            + core::ops::Add<Output = T>
+            + core::ops::Sub<Output = T>
+            + core::ops::Mul<Output = T>
+            + core::ops::Div<Output = T>
+            + Copy,
+    {
+        let overlap = self.overlap(other);
+        let union = self.area() + other.area() - overlap;
+        overlap / union
+    }
+
+    pub fn contains(&self, point: Point<T>) -> bool
+    where
+        T: std::cmp::PartialOrd + core::ops::Add<Output = T> + Copy,
+    {
+        point.x >= self.x
+            && point.x <= self.x + self.width
+            && point.y >= self.y
+            && point.y <= self.y + self.height
+    }
+
+    pub fn contains_bbox(&self, other: Self) -> bool
+    where
+        T: std::cmp::PartialOrd + Copy,
+        T: core::ops::Add<Output = T>,
+    {
+        self.contains(other.top_left())
+            && self.contains(other.top_right())
+            && self.contains(other.bottom_left())
+            && self.contains(other.bottom_right())
+    }
+
+    pub fn new_xywh(x: T, y: T, width: T, height: T) -> Self {
+        Self {
+            x,
+            y,
+            width,
+            height,
+        }
+    }
+    pub fn new_xyxy(x1: T, y1: T, x2: T, y2: T) -> Self
+    where
+        T: core::ops::Sub<Output = T> + Copy,
+    {
+        Self {
+            x: x1,
+            y: y1,
+            width: x2 - x1,
+            height: y2 - y1,
+        }
+    }
+
+    pub fn containing(box1: Self, box2: Self) -> Self
+    where
+        T: traits::min::Min + traits::max::Max + Copy,
+        T: core::ops::Sub<Output = T>,
+        T: core::ops::Add<Output = T>,
+    {
+        let x1 = box1.x.min(box2.x);
+        let y1 = box1.y.min(box2.y);
+        let x2 = box1.x2().max(box2.x2());
+        let y2 = box1.y2().max(box2.y2());
+        Self::new_xyxy(x1, y1, x2, y2)
+    }
+}
+
+impl<T: core::ops::Sub<Output = T> + Copy> core::ops::Sub<T> for BBox<T> {
+    type Output = BBox<T>;
+    fn sub(self, rhs: T) -> Self::Output {
+        BBox {
+            x: self.x - rhs,
+            y: self.y - rhs,
+            width: self.width - rhs,
+            height: self.height - rhs,
+        }
+    }
+}
+
+impl<T: core::ops::Add<Output = T> + Copy> core::ops::Add<T> for BBox<T> {
+    type Output = BBox<T>;
+    fn add(self, rhs: T) -> Self::Output {
+        BBox {
+            x: self.x + rhs,
+            y: self.y + rhs,
+            width: self.width + rhs,
+            height: self.height + rhs,
+        }
+    }
+}
+impl<T: core::ops::Mul<Output = T> + Copy> core::ops::Mul<T> for BBox<T> {
+    type Output = BBox<T>;
+    fn mul(self, rhs: T) -> Self::Output {
+        BBox {
+            x: self.x * rhs,
+            y: self.y * rhs,
+            width: self.width * rhs,
+            height: self.height * rhs,
+        }
+    }
+}
+impl<T: core::ops::Div<Output = T> + Copy> core::ops::Div<T> for BBox<T> {
+    type Output = BBox<T>;
+    fn div(self, rhs: T) -> Self::Output {
+        BBox {
+            x: self.x / rhs,
+            y: self.y / rhs,
+            width: self.width / rhs,
+            height: self.height / rhs,
+        }
+    }
+}
+
+impl<T> core::ops::Add<BBox<T>> for BBox<T>
+where
+    T: core::ops::Sub<Output = T>
+        + core::ops::Add<Output = T>
+        + traits::min::Min
+        + traits::max::Max
+        + Copy,
+{
+    type Output = BBox<T>;
+    fn add(self, rhs: BBox<T>) -> Self::Output {
+        let x1 = self.x1().min(rhs.x1());
+        let y1 = self.y1().min(rhs.y1());
+        let x2 = self.x2().max(rhs.x2());
+        let y2 = self.y2().max(rhs.y2());
+        BBox::new_xyxy(x1, y1, x2, y2)
+    }
+}
+
+#[test]
+fn test_bbox_add() {
+    let bbox1: BBox<usize> = BBox::new_xyxy(0, 0, 10, 10);
+    let bbox2: BBox<usize> = BBox::new_xyxy(5, 5, 15, 15);
+    let bbox3: BBox<usize> = bbox1 + bbox2;
+    assert_eq!(bbox3, BBox::new_xyxy(0, 0, 15, 15).cast());
+}
+
+#[derive(
+    Debug, Copy, Clone, serde::Serialize, serde::Deserialize, PartialEq, PartialOrd, Eq, Ord, Hash,
+)]
+pub struct Point<T = f32> {
+    x: T,
+    y: T,
+}
+
+impl<T> Point<T> {
+    pub const fn new(x: T, y: T) -> Self {
+        Self { x, y }
+    }
+
+    pub const fn x(&self) -> T
+    where
+        T: Copy,
+    {
+        self.x
+    }
+
+    pub const fn y(&self) -> T
+    where
+        T: Copy,
+    {
+        self.y
+    }
+
+    pub fn cast<T2>(&self) -> Point<T2>
+    where
+        T: num::cast::AsPrimitive<T2>,
+        T2: Copy + 'static,
+    {
+        Point {
+            x: self.x.as_(),
+            y: self.y.as_(),
+        }
+    }
+}
+
+impl<T: core::ops::Sub<T, Output = T> + Copy> core::ops::Sub<Point<T>> for Point<T> {
+    type Output = Point<T>;
+    fn sub(self, rhs: Point<T>) -> Self::Output {
+        Point {
+            x: self.x - rhs.x,
+            y: self.y - rhs.y,
+        }
+    }
+}
+
+impl<T: core::ops::Add<T, Output = T> + Copy> core::ops::Add<Point<T>> for Point<T> {
+    type Output = Point<T>;
+    fn add(self, rhs: Point<T>) -> Self::Output {
+        Point {
+            x: self.x + rhs.x,
+            y: self.y + rhs.y,
+        }
+    }
+}
+
+impl<T: core::ops::Sub<Output = T> + Copy> Point<T> {
+    /// If both the boxes are in the same scale then make the translation of the origin to the
+    /// other box
+    pub fn with_origin(&self, origin: Self) -> Self {
+        *self - origin
+    }
+}
+
+impl<T: core::ops::Add<Output = T> + Copy> Point<T> {
+    pub fn translate(&self, point: Point<T>) -> Self {
+        *self + point
+    }
+}
+
+impl<I: num::Zero> BBox<I>
+where
+    I: num::cast::AsPrimitive<usize>,
+{
+    pub fn zeros_ndarray_2d<T: num::Zero + Copy>(&self) -> ndarray::Array2<T> {
+        ndarray::Array2::<T>::zeros((self.height.as_(), self.width.as_()))
+    }
+    pub fn zeros_ndarray_3d<T: num::Zero + Copy>(&self, channels: usize) -> ndarray::Array3<T> {
+        ndarray::Array3::<T>::zeros((self.height.as_(), self.width.as_(), channels))
+    }
+    pub fn ones_ndarray_2d<T: num::One + Copy>(&self) -> ndarray::Array2<T> {
+        ndarray::Array2::<T>::ones((self.height.as_(), self.width.as_()))
+    }
+}
+
+impl<T: num::Float> BBox<T> {
+    pub fn round(&self) -> Self {
+        Self {
+            x: self.x.round(),
+            y: self.y.round(),
+            width: self.width.round(),
+            height: self.height.round(),
+        }
+    }
+}
+
+#[cfg(test)]
+mod bbox_clamp_tests {
+    use super::*;
+    #[test]
+    pub fn bbox_test_clamp_box() {
+        let large_box = BBox::new(0, 0, 100, 100);
+        let small_box = BBox::new(10, 10, 20, 20);
+        let clamped = large_box.clamp_box(small_box);
+        assert_eq!(clamped, small_box);
+    }
+
+    #[test]
+    pub fn bbox_test_clamp_box_offset() {
+        let box_a = BBox::new(0, 0, 100, 100);
+        let box_b = BBox::new(-10, -10, 20, 20);
+        let clamped = box_b.clamp_box(box_a);
+        let expected = BBox::new(0, 0, 10, 10);
+        assert_eq!(expected, clamped);
+    }
+}
+
+#[cfg(test)]
+mod bbox_padding_tests {
+    use super::*;
+    #[test]
+    pub fn bbox_test_padding() {
+        let bbox = BBox::new(0, 0, 10, 10);
+        let padded = bbox.padding(2);
+        assert_eq!(padded, BBox::new(-2, -2, 14, 14));
+    }
+
+    #[test]
+    pub fn bbox_test_padding_height() {
+        let bbox = BBox::new(0, 0, 10, 10);
+        let padded = bbox.padding_height(2);
+        assert_eq!(padded, BBox::new(0, -2, 10, 14));
+    }
+
+    #[test]
+    pub fn bbox_test_padding_width() {
+        let bbox = BBox::new(0, 0, 10, 10);
+        let padded = bbox.padding_width(2);
+        assert_eq!(padded, BBox::new(-2, 0, 14, 10));
+    }
+
+    #[test]
+    pub fn bbox_test_clamped_padding() {
+        let bbox = BBox::new(0, 0, 10, 10);
+        let padded = bbox.padding(2);
+        let clamp = BBox::new(0, 0, 12, 12);
+        let clamped = padded.clamp_box(clamp);
+        assert_eq!(clamped, clamp);
+    }
+
+    #[test]
+    pub fn bbox_clamp_failure() {
+        let og = BBox::new(475.0, 79.625, 37.0, 282.15);
+        let padded = BBox {
+            x: 471.3,
+            y: 51.412499999999994,
+            width: 40.69999999999999,
+            height: 338.54999999999995,
+        };
+        let clamp = BBox::new(0.0, 0.0, 512.0, 512.0);
+        let sus = padded.clamp_box(clamp);
+        assert!(clamp.contains_bbox(sus));
+    }
+}
+
+#[cfg(test)]
+mod bbox_scale_tests {
+    use super::*;
+    #[test]
+    pub fn bbox_test_scale_int() {
+        let bbox = BBox::new(0, 0, 10, 10);
+        let scaled = bbox.scale(2);
+        assert_eq!(scaled, BBox::new(-5, -5, 20, 20));
+    }
+
+    #[test]
+    pub fn bbox_test_scale_float() {
+        let bbox = BBox::new(0, 0, 10, 10).cast();
+        let scaled = bbox.scale(1.05); // 5% increase
+        let l = 10.0 * 0.05;
+        assert_eq!(scaled, BBox::new(-l / 2.0, -l / 2.0, 10.0 + l, 10.0 + l));
+    }
+
+    #[test]
+    pub fn bbox_test_scale_float_negative() {
+        let bbox = BBox::new(0, 0, 10, 10).cast();
+        let scaled = bbox.scale(0.95); // 5% decrease
+        let l = -10.0 * 0.05;
+        assert_eq!(scaled, BBox::new(-l / 2.0, -l / 2.0, 10.0 + l, 10.0 + l));
+    }
+
+    #[test]
+    pub fn bbox_scale_float() {
+        let bbox = BBox::new_xywh(0, 0, 200, 200);
+        let scaled = bbox.cast::<f64>().scale(1.1).cast::<i32>().clamp(0, 1000);
+        let expected = BBox::new(0, 0, 220, 220);
+        assert_eq!(scaled, expected);
+    }
+    #[test]
+    pub fn add_padding_bbox_example() {
+        // let result = add_padding_bbox(
+        //     vec![Rect::new(100, 200, 300, 400)],
+        //     (0.1, 0.1),
+        //     (1000, 1000),
+        // );
+        //   assert_eq!(result[0], Rect::new(70, 160, 360, 480));
+        let bbox = BBox::new(100, 200, 300, 400);
+        let scaled = bbox.cast::<f64>().scale(1.2).cast::<i32>().clamp(0, 1000);
+        assert_eq!(bbox, BBox::new(100, 200, 300, 400));
+        assert_eq!(scaled, BBox::new(70, 160, 360, 480));
+    }
+    #[test]
+    pub fn scale_bboxes() {
+        // let result = scale_bboxes(Rect::new(100, 200, 300, 400), (1000, 1000), (500, 500));
+        // assert_eq!(result[0], Rect::new(200, 400, 600, 800));
+        let bbox = BBox::new(100, 200, 300, 400);
+        let scaled = bbox.scale(2);
+        assert_eq!(scaled, BBox::new(200, 400, 600, 800));
+    }
+}

bbox/src/traits.rs

@@ -0,0 +1,2 @@
+pub mod max;
+pub mod min;

bbox/src/traits/max.rs

@@ -0,0 +1,27 @@
+pub trait Max: Sized + Copy {
+    fn max(self, other: Self) -> Self;
+}
+
+macro_rules! impl_max {
+    ($($t:ty),*) => {
+        $(
+            impl Max for $t {
+                fn max(self, other: Self) -> Self {
+                    Ord::max(self, other)
+                }
+            }
+        )*
+    };
+    (float $($t:ty),*) => {
+        $(
+            impl Max for $t {
+                fn max(self, other: Self) -> Self {
+                    Self::max(self, other)
+                }
+            }
+        )*
+    };
+}
+
+impl_max!(usize, u8, u16, u32, u64, u128, isize, i8, i16, i32, i64, i128);
+impl_max!(float f32, f64);

bbox/src/traits/min.rs

@@ -0,0 +1,27 @@
+pub trait Min: Sized + Copy {
+    fn min(self, other: Self) -> Self;
+}
+
+macro_rules! impl_min {
+    ($($t:ty),*) => {
+        $(
+            impl Min for $t {
+                fn min(self, other: Self) -> Self {
+                    Ord::min(self, other)
+                }
+            }
+        )*
+    };
+    (float $($t:ty),*) => {
+        $(
+            impl Min for $t {
+                fn min(self, other: Self) -> Self {
+                    Self::min(self, other)
+                }
+            }
+        )*
+    };
+}
+
+impl_min!(usize, u8, u16, u32, u64, u128, isize, i8, i16, i32, i64, i128);
+impl_min!(float f32, f64);

bounding-box/Cargo.toml

@@ -6,12 +6,16 @@ edition = "2024"
 [dependencies]
 color = "0.3.1"
 itertools = "0.14.0"
-nalgebra = "0.33.2"
+nalgebra = { workspace = true }
 ndarray = { version = "0.16.1", optional = true }
 num = "0.4.3"
+ordered-float = "5.0.0"
 simba = "0.9.0"
 thiserror = "2.0.12"
+tracing = { version = "0.1.41", optional = true, default-features = false }
 
 [features]
 ndarray = ["dep:ndarray"]
-default = ["ndarray"]
+tracing = ["dep:tracing"]
+
+default = ["ndarray", "tracing"]

bounding-box/src/draw.rs

@@ -4,11 +4,11 @@ pub use color::Rgba8;
 use ndarray::{Array1, Array3, ArrayViewMut3};
 
 pub trait Draw<T> {
-    fn draw(&mut self, item: T, color: color::Rgba8, thickness: usize);
+    fn draw(&mut self, item: &T, color: color::Rgba8, thickness: usize);
 }
 
 impl Draw<Aabb2<usize>> for Array3<u8> {
-    fn draw(&mut self, item: Aabb2<usize>, color: color::Rgba8, thickness: usize) {
+    fn draw(&mut self, item: &Aabb2<usize>, color: color::Rgba8, thickness: usize) {
         item.draw(self, color, thickness)
     }
 }
@@ -65,8 +65,9 @@ impl Drawable<Array3<u8>> for Aabb2<usize> {
                             pixel.assign(&color);
                         })
                 })
-                .inspect_err(|e| {
-                    dbg!(e);
+                .inspect_err(|_e| {
+                    #[cfg(feature = "tracing")]
+                    tracing::error!("{_e}")
                 })
                 .ok();
         });

bounding-box/src/lib.rs

@@ -2,9 +2,38 @@ pub mod draw;
 pub mod nms;
 pub mod roi;
 
-use nalgebra::{Point, Point2, Point3, SVector, SimdPartialOrd, SimdValue};
-pub trait Num: num::Num + Copy + core::fmt::Debug + 'static {}
-impl<T: num::Num + Copy + core::fmt::Debug + 'static> Num for T {}
+use nalgebra::{Point, Point2, SVector, Vector2};
+pub trait Num:
+    num::Num
+    + core::ops::AddAssign
+    + core::ops::SubAssign
+    + core::ops::MulAssign
+    + core::ops::DivAssign
+    + core::cmp::PartialOrd
+    + core::cmp::PartialEq
+    + nalgebra::SimdPartialOrd
+    + nalgebra::SimdValue
+    + Copy
+    + core::fmt::Debug
+    + 'static
+{
+}
+impl<
+    T: num::Num
+        + core::ops::AddAssign
+        + core::ops::SubAssign
+        + core::ops::MulAssign
+        + core::ops::DivAssign
+        + core::cmp::PartialOrd
+        + core::cmp::PartialEq
+        + nalgebra::SimdPartialOrd
+        + nalgebra::SimdValue
+        + Copy
+        + core::fmt::Debug
+        + 'static,
+> Num for T
+{
+}
 
 /// An axis aligned bounding box in `D` dimensions, defined by the minimum vertex and a size vector.
 #[derive(Debug, Copy, Clone, PartialEq, Eq)]
@@ -20,16 +49,27 @@ pub type Aabb2<T> = AxisAlignedBoundingBox<T, 2>;
 pub type Aabb3<T> = AxisAlignedBoundingBox<T, 3>;
 
 impl<T: Num, const D: usize> AxisAlignedBoundingBox<T, D> {
-    pub fn new(point: Point<T, D>, size: SVector<T, D>) -> Self {
+    // Panics if max < min
+    pub fn new(min_point: Point<T, D>, max_point: Point<T, D>) -> Self {
+        if max_point >= min_point {
+            Self::from_min_max_vertices(min_point, max_point)
+        } else {
+            panic!("max_point must be greater than or equal to min_point");
+        }
+    }
+    pub fn try_new(min_point: Point<T, D>, max_point: Point<T, D>) -> Option<Self> {
+        if max_point < min_point {
+            return None;
+        }
+        Some(Self::from_min_max_vertices(min_point, max_point))
+    }
+    pub fn new_point_size(point: Point<T, D>, size: SVector<T, D>) -> Self {
         Self { point, size }
     }
 
-    pub fn from_min_max_vertices(point1: Point<T, D>, point2: Point<T, D>) -> Self
-    where
-        T: core::ops::SubAssign,
-    {
-        let size = point2 - point1;
-        Self::new(point1, SVector::from(size))
+    pub fn from_min_max_vertices(min: Point<T, D>, max: Point<T, D>) -> Self {
+        let size = max - min;
+        Self::new_point_size(min, SVector::from(size))
     }
 
     /// Only considers the points closest and furthest from origin
@@ -123,6 +163,21 @@ impl<T: Num, const D: usize> AxisAlignedBoundingBox<T, D> {
         }
     }
 
+    pub fn scale_uniform(self, scalar: T) -> Self
+    where
+        T: core::ops::MulAssign,
+        T: core::ops::DivAssign,
+        T: core::ops::SubAssign,
+    {
+        let two = T::one() + T::one();
+        let new_size = self.size * scalar;
+        let new_point = self.point.coords - (new_size - self.size) / two;
+        Self {
+            point: Point::from(new_point),
+            size: new_size,
+        }
+    }
+
     pub fn contains_bbox(&self, other: &Self) -> bool
     where
         T: core::ops::AddAssign,
@@ -151,7 +206,21 @@ impl<T: Num, const D: usize> AxisAlignedBoundingBox<T, D> {
         self.intersection(other)
     }
 
-    pub fn union(&self, other: &Self) -> Self
+    pub fn component_clamp(&self, min: T, max: T) -> Self
+    where
+        T: PartialOrd,
+    {
+        let mut this = *self;
+        this.point.iter_mut().for_each(|x| {
+            *x = nalgebra::clamp(*x, min, max);
+        });
+        this.size.iter_mut().for_each(|x| {
+            *x = nalgebra::clamp(*x, min, max);
+        });
+        this
+    }
+
+    pub fn merge(&self, other: &Self) -> Self
     where
         T: core::ops::AddAssign,
         T: core::ops::SubAssign,
@@ -159,13 +228,24 @@ impl<T: Num, const D: usize> AxisAlignedBoundingBox<T, D> {
         T: nalgebra::SimdValue,
         T: nalgebra::SimdPartialOrd,
     {
-        let self_min = self.min_vertex();
-        let self_max = self.max_vertex();
-        let other_min = other.min_vertex();
-        let other_max = other.max_vertex();
-        let min = self_min.inf(&other_min);
-        let max = self_max.sup(&other_max);
-        Self::from_min_max_vertices(min, max)
+        let min = self.min_vertex().inf(&other.min_vertex());
+        let max = self.min_vertex().sup(&other.max_vertex());
+        Self::new(min, max)
+    }
+
+    pub fn union(&self, other: &Self) -> T
+    where
+        T: core::ops::AddAssign,
+        T: core::ops::SubAssign,
+        T: core::ops::MulAssign,
+        T: PartialOrd,
+        T: nalgebra::SimdValue,
+        T: nalgebra::SimdPartialOrd,
+    {
+        self.measure() + other.measure()
+            - Self::intersection(self, other)
+                .map(|x| x.measure())
+                .unwrap_or(T::zero())
     }
 
     pub fn intersection(&self, other: &Self) -> Option<Self>
@@ -176,21 +256,9 @@ impl<T: Num, const D: usize> AxisAlignedBoundingBox<T, D> {
         T: nalgebra::SimdPartialOrd,
         T: nalgebra::SimdValue,
     {
-        let self_min = self.min_vertex();
-        let self_max = self.max_vertex();
-        let other_min = other.min_vertex();
-        let other_max = other.max_vertex();
-
-        if self_max < other_min || other_max < self_min {
-            return None; // No intersection
-        }
-
-        let min = self_min.sup(&other_min);
-        let max = self_max.inf(&other_max);
-        Some(Self::from_min_max_vertices(
-            Point::from(min),
-            Point::from(max),
-        ))
+        let inter_min = self.min_vertex().sup(&other.min_vertex());
+        let inter_max = self.max_vertex().inf(&other.max_vertex());
+        Self::try_new(inter_min, inter_max)
     }
 
     pub fn denormalize(&self, factor: nalgebra::SVector<T, D>) -> Self
@@ -217,15 +285,17 @@ impl<T: Num, const D: usize> AxisAlignedBoundingBox<T, D> {
         })
     }
 
-    // pub fn as_<T2>(&self) -> Option<Aabb<T2, D>>
-    // where
-    //     T2: Num + simba::scalar::SubsetOf<T>,
-    // {
-    //     Some(Aabb {
-    //         point: Point::from(self.point.coords.as_()),
-    //         size: self.size.as_(),
-    //     })
-    // }
+    pub fn as_<T2>(&self) -> Aabb<T2, D>
+    where
+        T2: Num,
+        T: num::cast::AsPrimitive<T2>,
+    {
+        Aabb {
+            point: Point::from(self.point.coords.map(|x| x.as_())),
+            size: self.size.map(|x| x.as_()),
+        }
+    }
+
     pub fn measure(&self) -> T
     where
         T: core::ops::MulAssign,
@@ -233,7 +303,7 @@ impl<T: Num, const D: usize> AxisAlignedBoundingBox<T, D> {
         self.size.product()
     }
 
-    pub fn iou(&self, other: &Self) -> Option<T>
+    pub fn iou(&self, other: &Self) -> T
     where
         T: core::ops::AddAssign,
         T: core::ops::SubAssign,
@@ -242,9 +312,19 @@ impl<T: Num, const D: usize> AxisAlignedBoundingBox<T, D> {
         T: nalgebra::SimdValue,
         T: core::ops::MulAssign,
     {
-        let intersection = self.intersection(other)?;
-        let union = self.union(other);
-        Some(intersection.measure() / union.measure())
+        let lhs_min = self.min_vertex();
+        let lhs_max = self.max_vertex();
+        let rhs_min = other.min_vertex();
+        let rhs_max = other.max_vertex();
+
+        let inter_min = lhs_min.sup(&rhs_min);
+        let inter_max = lhs_max.inf(&rhs_max);
+        if inter_max >= inter_min {
+            let intersection = Aabb::new(inter_min, inter_max).measure();
+            intersection / (self.measure() + other.measure() - intersection)
+        } else {
+            return T::zero();
+        }
     }
 }
 
@@ -255,15 +335,15 @@ impl<T: Num> Aabb2<T> {
     {
         let point1 = Point2::new(x1, y1);
         let point2 = Point2::new(x2, y2);
-        Self::from_min_max_vertices(point1, point2)
+        Self::new(point1, point2)
     }
-    pub fn new_2d(point1: Point2<T>, point2: Point2<T>) -> Self
-    where
-        T: core::ops::SubAssign,
-    {
-        let size = point2.coords - point1.coords;
-        Self::new(point1, SVector::from(size))
+
+    pub fn from_xywh(x: T, y: T, w: T, h: T) -> Self {
+        let point = Point2::new(x, y);
+        let size = Vector2::new(w, h);
+        Self::new_point_size(point, size)
     }
+
     pub fn x1y1(&self) -> Point2<T> {
         self.point
     }
@@ -327,14 +407,6 @@ impl<T: Num> Aabb2<T> {
 }
 
 impl<T: Num> Aabb3<T> {
-    pub fn new_3d(point1: Point3<T>, point2: Point3<T>) -> Self
-    where
-        T: core::ops::SubAssign,
-    {
-        let size = point2.coords - point1.coords;
-        Self::new(point1, SVector::from(size))
-    }
-
     pub fn volume(&self) -> T
     where
         T: core::ops::MulAssign,
@@ -343,124 +415,125 @@ impl<T: Num> Aabb3<T> {
     }
 }
 
-#[test]
-fn test_bbox_new() {
-    use nalgebra::{Point2, Vector2};
+#[cfg(test)]
+mod boudning_box_tests {
+    use super::*;
+    use nalgebra::*;
 
+    #[test]
+    fn test_bbox_new() {
         let point1 = Point2::new(1.0, 2.0);
         let point2 = Point2::new(4.0, 6.0);
-    let bbox = AxisAlignedBoundingBox::new_2d(point1, point2);
+        let bbox = AxisAlignedBoundingBox::new(point1, point2);
 
         assert_eq!(bbox.min_vertex(), point1);
         assert_eq!(bbox.size(), Vector2::new(3.0, 4.0));
         assert_eq!(bbox.center(), Point2::new(2.5, 4.0));
-}
+    }
 
-#[test]
-fn test_bounding_box_center_2d() {
-    use nalgebra::{Point2, Vector2};
+    #[test]
+    fn test_intersection_and_merge() {
+        let point1 = Point2::new(1, 5);
+        let point2 = Point2::new(3, 2);
+        let size1 = Vector2::new(3, 4);
+        let size2 = Vector2::new(1, 3);
 
+        let this = Aabb2::new_point_size(point1, size1);
+        let other = Aabb2::new_point_size(point2, size2);
+        let inter = this.intersection(&other);
+        let merged = this.merge(&other);
+        assert_ne!(inter, Some(merged))
+    }
+
+    #[test]
+    fn test_bounding_box_center_2d() {
         let point = Point2::new(1.0, 2.0);
         let size = Vector2::new(3.0, 4.0);
-    let bbox = AxisAlignedBoundingBox::new(point, size);
+        let bbox = AxisAlignedBoundingBox::new_point_size(point, size);
 
         assert_eq!(bbox.min_vertex(), point);
         assert_eq!(bbox.size(), size);
         assert_eq!(bbox.center(), Point2::new(2.5, 4.0));
-}
-
-#[test]
-fn test_bounding_box_center_3d() {
-    use nalgebra::{Point3, Vector3};
+    }
 
+    #[test]
+    fn test_bounding_box_center_3d() {
         let point = Point3::new(1.0, 2.0, 3.0);
         let size = Vector3::new(4.0, 5.0, 6.0);
-    let bbox = AxisAlignedBoundingBox::new(point, size);
+        let bbox = AxisAlignedBoundingBox::new_point_size(point, size);
 
         assert_eq!(bbox.min_vertex(), point);
         assert_eq!(bbox.size(), size);
         assert_eq!(bbox.center(), Point3::new(3.0, 4.5, 6.0));
-}
-
-#[test]
-fn test_bounding_box_padding_2d() {
-    use nalgebra::{Point2, Vector2};
+    }
 
+    #[test]
+    fn test_bounding_box_padding_2d() {
         let point = Point2::new(1.0, 2.0);
         let size = Vector2::new(3.0, 4.0);
-    let bbox = AxisAlignedBoundingBox::new(point, size);
+        let bbox = AxisAlignedBoundingBox::new_point_size(point, size);
 
         let padded_bbox = bbox.padding(1.0);
         assert_eq!(padded_bbox.min_vertex(), Point2::new(0.5, 1.5));
         assert_eq!(padded_bbox.size(), Vector2::new(4.0, 5.0));
-}
-
-#[test]
-fn test_bounding_box_scaling_2d() {
-    use nalgebra::{Point2, Vector2};
+    }
 
+    #[test]
+    fn test_bounding_box_scaling_2d() {
         let point = Point2::new(1.0, 1.0);
         let size = Vector2::new(3.0, 4.0);
-    let bbox = AxisAlignedBoundingBox::new(point, size);
+        let bbox = AxisAlignedBoundingBox::new_point_size(point, size);
 
         let padded_bbox = bbox.scale(Vector2::new(2.0, 2.0));
         assert_eq!(padded_bbox.min_vertex(), Point2::new(-2.0, -3.0));
         assert_eq!(padded_bbox.size(), Vector2::new(6.0, 8.0));
-}
-
-#[test]
-fn test_bounding_box_contains_2d() {
-    use nalgebra::Point2;
+    }
 
+    #[test]
+    fn test_bounding_box_contains_2d() {
         let point1 = Point2::new(1.0, 2.0);
         let point2 = Point2::new(4.0, 6.0);
-    let bbox = AxisAlignedBoundingBox::new_2d(point1, point2);
+        let bbox = AxisAlignedBoundingBox::new(point1, point2);
 
         assert!(bbox.contains_point(&Point2::new(2.0, 3.0)));
         assert!(!bbox.contains_point(&Point2::new(5.0, 7.0)));
-}
-
-#[test]
-fn test_bounding_box_union_2d() {
-    use nalgebra::{Point2, Vector2};
+    }
 
+    #[test]
+    fn test_bounding_box_union_2d() {
         let point1 = Point2::new(1.0, 2.0);
         let point2 = Point2::new(4.0, 6.0);
-    let bbox1 = AxisAlignedBoundingBox::new_2d(point1, point2);
+        let bbox1 = AxisAlignedBoundingBox::new(point1, point2);
 
         let point3 = Point2::new(3.0, 5.0);
         let point4 = Point2::new(7.0, 8.0);
-    let bbox2 = AxisAlignedBoundingBox::new_2d(point3, point4);
+        let bbox2 = AxisAlignedBoundingBox::new(point3, point4);
 
-    let union_bbox = bbox1.union(&bbox2);
+        let union_bbox = bbox1.merge(&bbox2);
         assert_eq!(union_bbox.min_vertex(), Point2::new(1.0, 2.0));
         assert_eq!(union_bbox.size(), Vector2::new(6.0, 6.0));
-}
-
-#[test]
-fn test_bounding_box_intersection_2d() {
-    use nalgebra::{Point2, Vector2};
+    }
 
+    #[test]
+    fn test_bounding_box_intersection_2d() {
         let point1 = Point2::new(1.0, 2.0);
         let point2 = Point2::new(4.0, 6.0);
-    let bbox1 = AxisAlignedBoundingBox::new_2d(point1, point2);
+        let bbox1 = AxisAlignedBoundingBox::new(point1, point2);
 
         let point3 = Point2::new(3.0, 5.0);
         let point4 = Point2::new(5.0, 7.0);
-    let bbox2 = AxisAlignedBoundingBox::new_2d(point3, point4);
+        let bbox2 = AxisAlignedBoundingBox::new(point3, point4);
 
         let intersection_bbox = bbox1.intersection(&bbox2).unwrap();
         assert_eq!(intersection_bbox.min_vertex(), Point2::new(3.0, 5.0));
         assert_eq!(intersection_bbox.size(), Vector2::new(1.0, 1.0));
-}
-
-#[test]
-fn test_bounding_box_contains_point() {
-    use nalgebra::Point2;
+    }
 
+    #[test]
+    fn test_bounding_box_contains_point() {
         let point1 = Point2::new(2, 3);
         let point2 = Point2::new(5, 4);
-    let bbox = AxisAlignedBoundingBox::new_2d(point1, point2);
+        let bbox = AxisAlignedBoundingBox::new(point1, point2);
         use itertools::Itertools;
         for (i, j) in (0..=10).cartesian_product(0..=10) {
             if bbox.contains_point(&Point2::new(i, j)) {
@@ -479,10 +552,10 @@ fn test_bounding_box_contains_point() {
                 }
             }
         }
-}
+    }
 
-#[test]
-fn test_bounding_box_clamp_box_2d() {
+    #[test]
+    fn test_bounding_box_clamp_box_2d() {
         let bbox1 = Aabb2::from_x1y1x2y2(1, 1, 4, 4);
         let bbox2 = Aabb2::from_x1y1x2y2(2, 2, 3, 3);
         let clamped = bbox2.clamp(&bbox1).unwrap();
@@ -495,4 +568,63 @@ fn test_bounding_box_clamp_box_2d() {
         let clamped = bbox1.clamp(&bbox2).unwrap();
         let expected = Aabb2::from_x1y1x2y2(5, 5, 7, 7);
         assert_eq!(clamped, expected)
+    }
+
+    #[test]
+    fn test_iou_identical_boxes() {
+        let a = Aabb2::from_x1y1x2y2(1.0, 2.0, 4.0, 6.0);
+        let b = Aabb2::from_x1y1x2y2(1.0, 2.0, 4.0, 6.0);
+        assert_eq!(a.iou(&b), 1.0);
+    }
+
+    #[test]
+    fn test_iou_non_overlapping_boxes() {
+        let a = Aabb2::from_x1y1x2y2(0.0, 0.0, 1.0, 1.0);
+        let b = Aabb2::from_x1y1x2y2(2.0, 2.0, 3.0, 3.0);
+        assert_eq!(a.iou(&b), 0.0);
+    }
+
+    #[test]
+    fn test_iou_partial_overlap() {
+        let a = Aabb2::from_x1y1x2y2(0.0, 0.0, 2.0, 2.0);
+        let b = Aabb2::from_x1y1x2y2(1.0, 1.0, 3.0, 3.0);
+        // Intersection area = 1, Union area = 7
+        assert!((a.iou(&b) - 1.0 / 7.0).abs() < 1e-6);
+    }
+
+    #[test]
+    fn test_iou_one_inside_another() {
+        let a = Aabb2::from_x1y1x2y2(0.0, 0.0, 4.0, 4.0);
+        let b = Aabb2::from_x1y1x2y2(1.0, 1.0, 3.0, 3.0);
+        // Intersection area = 4, Union area = 16
+        assert!((a.iou(&b) - 0.25).abs() < 1e-6);
+    }
+
+    #[test]
+    fn test_iou_edge_touching() {
+        let a = Aabb2::from_x1y1x2y2(0.0, 0.0, 1.0, 1.0);
+        let b = Aabb2::from_x1y1x2y2(1.0, 0.0, 2.0, 1.0);
+        assert_eq!(a.iou(&b), 0.0);
+    }
+
+    #[test]
+    fn test_iou_corner_touching() {
+        let a = Aabb2::from_x1y1x2y2(0.0, 0.0, 1.0, 1.0);
+        let b = Aabb2::from_x1y1x2y2(1.0, 1.0, 2.0, 2.0);
+        assert_eq!(a.iou(&b), 0.0);
+    }
+
+    #[test]
+    fn test_iou_zero_area_box() {
+        let a = Aabb2::from_x1y1x2y2(0.0, 0.0, 0.0, 0.0);
+        let b = Aabb2::from_x1y1x2y2(0.0, 0.0, 1.0, 1.0);
+        assert_eq!(a.iou(&b), 0.0);
+    }
+
+    #[test]
+    fn test_specific_values() {
+        let box1 = Aabb2::from_xywh(0.69482, 0.6716774, 0.07493961, 0.14968264);
+        let box2 = Aabb2::from_xywh(0.41546485, 0.70290875, 0.06197411, 0.08818436);
+        assert!(box1.iou(&box2) >= 0.0);
+    }
 }

bounding-box/src/nms.rs

@@ -1,4 +1,11 @@
-use std::collections::HashSet;
+use std::collections::{HashSet, VecDeque};
+
+use itertools::Itertools;
+#[derive(Debug, Clone, Copy, PartialEq, Eq, thiserror::Error)]
+pub enum NmsError {
+    #[error("Boxes and scores length mismatch (boxes: {boxes}, scores: {scores})")]
+    BoxesAndScoresLengthMismatch { boxes: usize, scores: usize },
+}
 
 use crate::*;
 /// Apply Non-Maximum Suppression to a set of bounding boxes.
@@ -18,10 +25,11 @@ pub fn nms<T>(
     scores: &[T],
     score_threshold: T,
     nms_threshold: T,
-) -> HashSet<usize>
+) -> Result<HashSet<usize>, NmsError>
 where
     T: Num
-        + num::Float
+        + ordered_float::FloatCore
+        + core::ops::Neg<Output = T>
         + core::iter::Product<T>
         + core::ops::AddAssign
         + core::ops::SubAssign
@@ -29,56 +37,37 @@ where
         + nalgebra::SimdValue
         + nalgebra::SimdPartialOrd,
 {
-    use itertools::Itertools;
-
-    // Create vector of (index, box, score) tuples for boxes with scores above threshold
-    let mut indexed_boxes: Vec<(usize, &Aabb2<T>, &T)> = boxes
+    if boxes.len() != scores.len() {
+        return Err(NmsError::BoxesAndScoresLengthMismatch {
+            boxes: boxes.len(),
+            scores: scores.len(),
+        });
+    }
+    let mut combined: VecDeque<(usize, Aabb2<T>, T, bool)> = boxes
         .iter()
         .enumerate()
-        .zip(scores.iter())
+        .zip(scores)
         .filter_map(|((idx, bbox), score)| {
-            if *score >= score_threshold {
-                Some((idx, bbox, score))
-            } else {
-                None
-            }
+            (*score > score_threshold).then_some((idx, *bbox, *score, true))
         })
+        .sorted_by_cached_key(|(_, _, score, _)| -ordered_float::OrderedFloat(*score))
         .collect();
 
-    // Sort by score in descending order
-    indexed_boxes.sort_by(|(_, _, score_a), (_, _, score_b)| {
-        score_b
-            .partial_cmp(score_a)
-            .unwrap_or(std::cmp::Ordering::Equal)
-    });
-
-    let mut keep_indices = HashSet::new();
-    let mut suppressed = HashSet::new();
-
-    for (i, (idx_i, bbox_i, _)) in indexed_boxes.iter().enumerate() {
-        // Skip if this box is already suppressed
-        if suppressed.contains(idx_i) {
+    for i in 0..combined.len() {
+        let first = combined[i];
+        if first.3 == false {
             continue;
         }
-
-        // Keep this box
-        keep_indices.insert(*idx_i);
-
-        // Compare with remaining boxes
-        for (idx_j, bbox_j, _) in indexed_boxes.iter().skip(i + 1) {
-            // Skip if this box is already suppressed
-            if suppressed.contains(idx_j) {
-                continue;
-            }
-
-            // Calculate IoU and suppress if above threshold
-            if let Some(iou) = bbox_i.iou(bbox_j) {
-                if iou >= nms_threshold {
-                    suppressed.insert(*idx_j);
-                }
+        let bbox = first.1;
+        for item in combined.iter_mut().skip(i + 1) {
+            if bbox.iou(&item.1) > nms_threshold {
+                item.3 = false
             }
         }
     }
 
-    keep_indices
+    Ok(combined
+        .into_iter()
+        .filter_map(|(idx, _, _, keep)| keep.then_some(idx))
+        .collect())
 }

bounding-box/src/roi.rs

@@ -5,10 +5,17 @@ pub trait Roi<'a, Output> {
     type Error;
     fn roi(&'a self, aabb: Aabb2<usize>) -> Result<Output, Self::Error>;
 }
+
 pub trait RoiMut<'a, Output> {
     type Error;
     fn roi_mut(&'a mut self, aabb: Aabb2<usize>) -> Result<Output, Self::Error>;
 }
+
+pub trait MultiRoi<'a, Output> {
+    type Error;
+    fn multi_roi(&'a self, aabbs: &[Aabb2<usize>]) -> Result<Output, Self::Error>;
+}
+
 #[derive(thiserror::Error, Debug, Copy, Clone)]
 pub enum RoiError {
     #[error("Region of intereset is out of bounds")]
@@ -36,7 +43,7 @@ impl<'a, T: Num> RoiMut<'a, ArrayViewMut3<'a, T>> for Array3<T> {
         let x2 = aabb.x2();
         let y1 = aabb.y1();
         let y2 = aabb.y2();
-        if x1 >= x2 || y1 >= y2 || x2 > self.shape()[1] || y2 > self.shape()[0] {
+        if x1 > x2 || y1 > y2 || x2 > self.shape()[1] || y2 > self.shape()[0] {
             return Err(RoiError::RoiOutOfBounds);
         }
         Ok(self.slice_mut(ndarray::s![y1..y2, x1..x2, ..]))
@@ -95,3 +102,47 @@ pub fn reborrow_test() {
     };
     dbg!(y);
 }
+
+impl<'a> MultiRoi<'a, Vec<ArrayView3<'a, u8>>> for Array3<u8> {
+    type Error = RoiError;
+    fn multi_roi(&'a self, aabbs: &[Aabb2<usize>]) -> Result<Vec<ArrayView3<'a, u8>>, Self::Error> {
+        let (height, width, _channels) = self.dim();
+        let outer_aabb = Aabb2::from_x1y1x2y2(0, 0, width, height);
+        aabbs
+            .iter()
+            .map(|aabb| {
+                let slice_arg =
+                    bbox_to_slice_arg(aabb.clamp(&outer_aabb).ok_or(RoiError::RoiOutOfBounds)?);
+                Ok(self.slice(slice_arg))
+            })
+            .collect::<Result<Vec<_>, RoiError>>()
+    }
+}
+
+impl<'a, 'b> MultiRoi<'a, Vec<ArrayView3<'b, u8>>> for ArrayView3<'b, u8> {
+    type Error = RoiError;
+    fn multi_roi(&'a self, aabbs: &[Aabb2<usize>]) -> Result<Vec<ArrayView3<'b, u8>>, Self::Error> {
+        let (height, width, _channels) = self.dim();
+        let outer_aabb = Aabb2::from_x1y1x2y2(0, 0, width, height);
+        aabbs
+            .iter()
+            .map(|aabb| {
+                let slice_arg =
+                    bbox_to_slice_arg(aabb.clamp(&outer_aabb).ok_or(RoiError::RoiOutOfBounds)?);
+                Ok(self.slice_move(slice_arg))
+            })
+            .collect::<Result<Vec<_>, RoiError>>()
+    }
+}
+
+fn bbox_to_slice_arg(
+    aabb: Aabb2<usize>,
+) -> ndarray::SliceInfo<[ndarray::SliceInfoElem; 3], ndarray::Ix3, ndarray::Ix3> {
+    // This function should convert the bounding box to a slice argument
+    // For now, we will return a dummy value
+    let x1 = aabb.x1();
+    let x2 = aabb.x2();
+    let y1 = aabb.y1();
+    let y2 = aabb.y2();
+    ndarray::s![y1..y2, x1..x2, ..]
+}

flake.lock

@@ -3,11 +3,11 @@
     "advisory-db": {
       "flake": false,
       "locked": {
-        "lastModified": 1750151065,
-        "narHash": "sha256-il+CAqChFIB82xP6bO43dWlUVs+NlG7a4g8liIP5HcI=",
+        "lastModified": 1755283329,
+        "narHash": "sha256-33bd+PHbon+cgEiWE/zkr7dpEF5E0DiHOzyoUQbkYBc=",
         "owner": "rustsec",
         "repo": "advisory-db",
-        "rev": "7573f55ba337263f61167dbb0ea926cdc7c8eb5d",
+        "rev": "61aac2116c8cb7cc80ff8ca283eec7687d384038",
         "type": "github"
       },
       "original": {
@@ -18,11 +18,11 @@
     },
     "crane": {
       "locked": {
-        "lastModified": 1750266157,
-        "narHash": "sha256-tL42YoNg9y30u7zAqtoGDNdTyXTi8EALDeCB13FtbQA=",
+        "lastModified": 1754269165,
+        "narHash": "sha256-0tcS8FHd4QjbCVoxN9jI+PjHgA4vc/IjkUSp+N3zy0U=",
         "owner": "ipetkov",
         "repo": "crane",
-        "rev": "e37c943371b73ed87faf33f7583860f81f1d5a48",
+        "rev": "444e81206df3f7d92780680e45858e31d2f07a08",
         "type": "github"
       },
       "original": {
@@ -109,16 +109,16 @@
     "mnn-src": {
       "flake": false,
       "locked": {
-        "lastModified": 1749173738,
-        "narHash": "sha256-pNljvQ4xMZ4VmuxQyXt+boNBZD0+UZNpNLrWrj8Rtfw=",
+        "lastModified": 1753256753,
+        "narHash": "sha256-aTpwVZBkpQiwOVVXDfQIVEx9CswNiPbvNftw8KsoW+Q=",
         "owner": "alibaba",
         "repo": "MNN",
-        "rev": "ebdada82634300956e08bd4056ecfeb1e4f23b32",
+        "rev": "a739ea5870a4a45680f0e36ba9662ca39f2f4eec",
         "type": "github"
       },
       "original": {
         "owner": "alibaba",
-        "ref": "3.2.0",
+        "ref": "3.2.2",
         "repo": "MNN",
         "type": "github"
       }
@@ -145,11 +145,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1750506804,
-        "narHash": "sha256-VLFNc4egNjovYVxDGyBYTrvVCgDYgENp5bVi9fPTDYc=",
+        "lastModified": 1755186698,
+        "narHash": "sha256-wNO3+Ks2jZJ4nTHMuks+cxAiVBGNuEBXsT29Bz6HASo=",
         "owner": "nixos",
         "repo": "nixpkgs",
-        "rev": "4206c4cb56751df534751b058295ea61357bbbaa",
+        "rev": "fbcf476f790d8a217c3eab4e12033dc4a0f6d23c",
         "type": "github"
       },
       "original": {
@@ -178,11 +178,11 @@
         ]
       },
       "locked": {
-        "lastModified": 1750732748,
-        "narHash": "sha256-HR2b3RHsPeJm+Fb+1ui8nXibgniVj7hBNvUbXEyz0DU=",
+        "lastModified": 1755485198,
+        "narHash": "sha256-C3042ST2lUg0nh734gmuP4lRRIBitA6Maegg2/jYRM4=",
         "owner": "oxalica",
         "repo": "rust-overlay",
-        "rev": "4b4494b2ba7e8a8041b2e28320b2ee02c115c75f",
+        "rev": "aa45e63d431b28802ca4490cfc796b9e31731df7",
         "type": "github"
       },
       "original": {

flake.nix

@@ -22,7 +22,7 @@
       inputs.nixpkgs.follows = "nixpkgs";
     };
     mnn-src = {
-      url = "github:alibaba/MNN/3.2.0";
+      url = "github:alibaba/MNN/3.2.2";
       flake = false;
     };
   };
@@ -43,13 +43,15 @@
       system: let
         pkgs = import nixpkgs {
           inherit system;
+          config.allowUnfree = true;
+          config.cudaSupport = pkgs.stdenv.isLinux;
           overlays = [
             rust-overlay.overlays.default
             (final: prev: {
               mnn = mnn-overlay.packages.${system}.mnn.override {
                 src = mnn-src;
                 buildConverter = true;
-                enableMetal = true;
+                enableMetal = pkgs.stdenv.isDarwin;
                 enableOpencl = true;
               };
             })
@@ -61,17 +63,44 @@
 
         stableToolchain = pkgs.rust-bin.stable.latest.default;
         stableToolchainWithLLvmTools = stableToolchain.override {
-          extensions = ["rust-src" "llvm-tools"];
+          extensions = [
+            "rust-src"
+            "llvm-tools"
+          ];
         };
         stableToolchainWithRustAnalyzer = stableToolchain.override {
-          extensions = ["rust-src" "rust-analyzer"];
+          extensions = [
+            "rust-src"
+            "rust-analyzer"
+          ];
         };
         craneLib = (crane.mkLib pkgs).overrideToolchain stableToolchain;
         craneLibLLvmTools = (crane.mkLib pkgs).overrideToolchain stableToolchainWithLLvmTools;
 
+        ort_static = (pkgs.onnxruntime.overide {cudaSupport = true;}).overrideAttrs (old: {
+          cmakeFlags =
+            old.cmakeFlags
+            ++ [
+              "-Donnxruntime_BUILD_SHARED_LIB=OFF"
+              "-Donnxruntime_BUILD_STATIC_LIB=ON"
+            ];
+        });
+        patchedOnnxruntime = pkgs.onnxruntime.overrideAttrs (old: {
+          patches = [./patches/ort_env_global_mutex.patch];
+        });
         src = let
           filterBySuffix = path: exts: lib.any (ext: lib.hasSuffix ext path) exts;
-          sourceFilters = path: type: (craneLib.filterCargoSources path type) || filterBySuffix path [".c" ".h" ".hpp" ".cpp" ".cc"];
+          sourceFilters = path: type:
+            (craneLib.filterCargoSources path type)
+            || filterBySuffix path [
+              ".c"
+              ".h"
+              ".hpp"
+              ".cpp"
+              ".cc"
+              ".mnn"
+              ".onnx"
+            ];
         in
           lib.cleanSourceWith {
             filter = sourceFilters;
@@ -81,15 +110,21 @@
           {
             inherit src;
             pname = name;
-            stdenv = pkgs.clangStdenv;
+            stdenv = p: p.clangStdenv;
             doCheck = false;
             LIBCLANG_PATH = "${pkgs.llvmPackages.libclang.lib}/lib";
-            # nativeBuildInputs = with pkgs; [
-            #   cmake
-            #   llvmPackages.libclang.lib
-            # ];
+            # ORT_LIB_LOCATION = "${patchedOnnxruntime}";
+            # ORT_ENV_SYSTEM_LIB_LOCATION = "${patchedOnnxruntime}/lib";
+            # ORT_ENV_PREFER_DYNAMIC_LINK = true;
+            nativeBuildInputs = with pkgs; [
+              cmake
+              pkg-config
+            ];
             buildInputs = with pkgs;
-              []
+              [
+                patchedOnnxruntime
+                sqlite
+              ]
               ++ (lib.optionals pkgs.stdenv.isDarwin [
                 libiconv
                 apple-sdk_13
@@ -102,11 +137,13 @@
       in {
         checks =
           {
-            "${name}-clippy" = craneLib.cargoClippy (commonArgs
+            "${name}-clippy" = craneLib.cargoClippy (
+              commonArgs
               // {
                 inherit cargoArtifacts;
                 cargoClippyExtraArgs = "--all-targets -- --deny warnings";
-              });
+              }
+            );
             "${name}-docs" = craneLib.cargoDoc (commonArgs // {inherit cargoArtifacts;});
             "${name}-fmt" = craneLib.cargoFmt {inherit src;};
             "${name}-toml-fmt" = craneLib.taploFmt {
@@ -121,22 +158,29 @@
             "${name}-deny" = craneLib.cargoDeny {
               inherit src;
             };
-            "${name}-nextest" = craneLib.cargoNextest (commonArgs
+            "${name}-nextest" = craneLib.cargoNextest (
+              commonArgs
               // {
                 inherit cargoArtifacts;
                 partitions = 1;
                 partitionType = "count";
-              });
+              }
+            );
           }
           // lib.optionalAttrs (!pkgs.stdenv.isDarwin) {
             "${name}-llvm-cov" = craneLibLLvmTools.cargoLlvmCov (commonArgs // {inherit cargoArtifacts;});
           };
 
         packages = let
-          pkg = craneLib.buildPackage (commonArgs
-            // {inherit cargoArtifacts;}
+          pkg = craneLib.buildPackage (
+            commonArgs
             // {
-              nativeBuildInputs = with pkgs; [
+              inherit cargoArtifacts;
+            }
+            // {
+              nativeBuildInputs = with pkgs;
+                commonArgs.nativeBuildInputs
+                ++ [
                   installShellFiles
                 ];
               postInstall = ''
@@ -145,28 +189,69 @@
                   --fish <($out/bin/${name} completions fish) \
                   --zsh <($out/bin/${name} completions zsh)
               '';
-            });
+            }
+          );
         in {
           "${name}" = pkg;
           default = pkg;
+          onnxruntime = ort_static;
         };
 
         devShells = {
-          default = pkgs.mkShell.override {stdenv = pkgs.clangStdenv;} (commonArgs
-            // {
+          default = pkgs.mkShell.override {stdenv = pkgs.clangStdenv;} (
+            commonArgs
+            // rec {
               LLDB_DEBUGSERVER_PATH = "/Applications/Xcode.app/Contents/SharedFrameworks/LLDB.framework/Versions/A/Resources/debugserver";
+              LD_LIBRARY_PATH = "$LD_LIBRARY_PATH:${builtins.toString (pkgs.lib.makeLibraryPath packages)}";
               packages = with pkgs;
                 [
                   stableToolchainWithRustAnalyzer
+                  cargo-expand
+                  cargo-outdated
                   cargo-nextest
                   cargo-deny
                   cmake
                   mnn
+                  cargo-make
+                  hyperfine
+                  opencv
+                  uv
+                  # (python312.withPackages (ps:
+                  #   with ps; [
+                  #     numpy
+                  #     matplotlib
+                  #     scikit-learn
+                  #     opencv-python
+                  #     seaborn
+                  #     torch
+                  #     torchvision
+                  #     tensorflow-lite
+                  #     retinaface
+                  #     facenet-pytorch
+                  #     tqdm
+                  #     pillow
+                  #     orjson
+                  #     huggingface-hub
+                  #     # insightface
+                  #   ]))
                 ]
                 ++ (lib.optionals pkgs.stdenv.isDarwin [
                   apple-sdk_13
+                ])
+                ++ (lib.optionals pkgs.stdenv.isLinux [
+                  xorg.libX11
+                  xorg.libXcursor
+                  xorg.libXrandr
+                  xorg.libXi
+                  xorg.libxcb
+                  libxkbcommon
+                  vulkan-loader
+                  wayland
+                  zenity
+                  cudatoolkit
                 ]);
-            });
+            }
+          );
         };
       }
     )

justfile

@@ -1,2 +1,13 @@
-run:
-    cargo run -r detect -- ./1000066593.jpg
+run_onnx ep = "cpu" arg = "selfie.jpg":
+    cargo run -r detect -p {{ep}} -t 0.3 -o detected.jpg -- {{arg}}
+run_mnn forward = "cpu" arg = "selfie.jpg":
+    cargo run -r detect -f {{forward}} -o detected.jpg -- {{arg}}
+
+open:
+    open detected.jpg
+
+bench:
+    cargo build --release
+    BINARY="" hyperfine --warmup 3 --export-markdown benchmark.md \
+        "$CARGO_TARGET_DIR/release/detector detect -f cpu selfie.jpg" \
+        "$CARGO_TARGET_DIR/release/detector detect -f cpu -b 1 selfie.jpg"

ndarray-image/src/lib.rs

@@ -5,7 +5,7 @@ fn shape_error() -> ndarray::ShapeError {
 
 mod rgb8 {
     use super::Result;
-    pub(super) fn image_as_ndarray(image: &image::RgbImage) -> Result<ndarray::ArrayView3<u8>> {
+    pub(super) fn image_as_ndarray(image: &image::RgbImage) -> Result<ndarray::ArrayView3<'_, u8>> {
         let (width, height) = image.dimensions();
         let data = image.as_raw();
         ndarray::ArrayView3::from_shape((height as usize, width as usize, 3), data)
@@ -31,7 +31,9 @@ mod rgb8 {
 
 mod rgba8 {
     use super::Result;
-    pub(super) fn image_as_ndarray(image: &image::RgbaImage) -> Result<ndarray::ArrayView3<u8>> {
+    pub(super) fn image_as_ndarray(
+        image: &image::RgbaImage,
+    ) -> Result<ndarray::ArrayView3<'_, u8>> {
         let (width, height) = image.dimensions();
         let data = image.as_raw();
         ndarray::ArrayView3::from_shape((height as usize, width as usize, 4), data)
@@ -57,7 +59,9 @@ mod rgba8 {
 
 mod gray8 {
     use super::Result;
-    pub(super) fn image_as_ndarray(image: &image::GrayImage) -> Result<ndarray::ArrayView2<u8>> {
+    pub(super) fn image_as_ndarray(
+        image: &image::GrayImage,
+    ) -> Result<ndarray::ArrayView2<'_, u8>> {
         let (width, height) = image.dimensions();
         let data = image.as_raw();
         ndarray::ArrayView2::from_shape((height as usize, width as usize), data)
@@ -82,7 +86,7 @@ mod gray_alpha8 {
     use super::Result;
     pub(super) fn image_as_ndarray(
         image: &image::GrayAlphaImage,
-    ) -> Result<ndarray::ArrayView3<u8>> {
+    ) -> Result<ndarray::ArrayView3<'_, u8>> {
         let (width, height) = image.dimensions();
         let data = image.as_raw();
         ndarray::ArrayView3::from_shape((height as usize, width as usize, 2), data)
@@ -110,7 +114,7 @@ mod gray_alpha8 {
 
 mod dynamic_image {
     use super::*;
-    pub fn image_as_ndarray(image: &image::DynamicImage) -> Result<ndarray::ArrayViewD<u8>> {
+    pub fn image_as_ndarray(image: &image::DynamicImage) -> Result<ndarray::ArrayViewD<'_, u8>> {
         Ok(match image {
             image::DynamicImage::ImageRgb8(img) => rgb8::image_as_ndarray(img)?.into_dyn(),
             image::DynamicImage::ImageRgba8(img) => rgba8::image_as_ndarray(img)?.into_dyn(),

ndarray-resize/src/lib.rs

@@ -147,7 +147,7 @@ impl<S: ndarray::Data<Elem = T>, T: seal::Sealed + bytemuck::Pod, D: ndarray::Di
     NdAsImage<T, D> for ndarray::ArrayBase<S, D>
 {
     /// Clones self and makes a new image
-    fn as_image_ref(&self) -> Result<ImageRef> {
+    fn as_image_ref(&self) -> Result<ImageRef<'_>> {
         let shape = self.shape();
         let rows = *shape
             .first()

ndarray-safetensors/Cargo.toml

@@ -0,0 +1,11 @@
+[package]
+name = "ndarray-safetensors"
+version.workspace = true
+edition.workspace = true
+
+[dependencies]
+bytemuck = { version = "1.23.2" }
+half = { version = "2.6.0", default-features = false, features = ["bytemuck"] }
+ndarray = { version = "0.16.1", default-features = false, features = ["std"] }
+safetensors = "0.6.2"
+thiserror = "2.0.15"

ndarray-safetensors/src/lib.rs

@@ -0,0 +1,449 @@
+//! # ndarray-serialize
+//!
+//! A Rust library for serializing and deserializing `ndarray` arrays using the SafeTensors format.
+//!
+//! ## Features
+//! - Serialize `ndarray::ArrayView` to SafeTensors format
+//! - Deserialize SafeTensors data back to `ndarray::ArrayView`
+//! - Support for multiple data types (f32, f64, i8-i64, u8-u64, f16, bf16)
+//! - Zero-copy deserialization when possible
+//! - Metadata support
+//!
+//! ## Example
+//! ```rust
+//! use ndarray::Array2;
+//! use ndarray_safetensors::{SafeArrays, SafeArrayView};
+//!
+//! // Create some data
+//! let array = Array2::<f32>::zeros((3, 4));
+//!
+//! // Serialize
+//! let mut safe_arrays = SafeArrays::new();
+//! safe_arrays.insert_ndarray("my_tensor", array.view()).unwrap();
+//! safe_arrays.insert_metadata("author", "example");
+//! let bytes = safe_arrays.serialize().unwrap();
+//!
+//! // Deserialize
+//! let view = SafeArrayView::from_bytes(&bytes).unwrap();
+//! let tensor: ndarray::ArrayView2<f32> = view.tensor("my_tensor").unwrap();
+//! assert_eq!(tensor.shape(), &[3, 4]);
+//! ```
+
+use safetensors::View;
+use std::borrow::Cow;
+use std::collections::{BTreeMap, HashMap};
+
+use thiserror::Error;
+/// Errors that can occur during SafeTensor operations
+#[derive(Error, Debug)]
+pub enum SafeTensorError {
+    #[error("Tensor not found: {0}")]
+    TensorNotFound(String),
+    #[error("Invalid tensor data: Got {0} Expected: {1}")]
+    InvalidTensorData(&'static str, String),
+    #[error("IO error: {0}")]
+    IoError(#[from] std::io::Error),
+    #[error("Safetensor error: {0}")]
+    SafeTensor(#[from] safetensors::SafeTensorError),
+    #[error("ndarray::ShapeError error: {0}")]
+    NdarrayShapeError(#[from] ndarray::ShapeError),
+}
+
+type Result<T, E = SafeTensorError> = core::result::Result<T, E>;
+
+use safetensors::tensor::SafeTensors;
+
+/// A view into SafeTensors data that provides access to ndarray tensors
+///
+/// # Example
+/// ```rust
+/// use ndarray::Array2;
+/// use ndarray_safetensors::{SafeArrays, SafeArrayView};
+///
+/// let array = Array2::<f32>::ones((2, 3));
+/// let mut safe_arrays = SafeArrays::new();
+/// safe_arrays.insert_ndarray("data", array.view()).unwrap();
+/// let bytes = safe_arrays.serialize().unwrap();
+///
+/// let view = SafeArrayView::from_bytes(&bytes).unwrap();
+/// let tensor: ndarray::ArrayView2<f32> = view.tensor("data").unwrap();
+/// ```
+#[derive(Debug)]
+pub struct SafeArraysView<'a> {
+    pub tensors: SafeTensors<'a>,
+}
+
+impl<'a> SafeArraysView<'a> {
+    fn new(tensors: SafeTensors<'a>) -> Self {
+        Self { tensors }
+    }
+
+    /// Create a SafeArrayView from serialized bytes
+    pub fn from_bytes(bytes: &'a [u8]) -> Result<SafeArraysView<'a>> {
+        let tensors = SafeTensors::deserialize(bytes)?;
+        Ok(Self::new(tensors))
+    }
+
+    /// Get a dynamic-dimensional tensor by name
+    pub fn dynamic_tensor<T: STDtype>(&self, name: &str) -> Result<ndarray::ArrayViewD<'a, T>> {
+        self.tensors
+            .tensor(name)
+            .map(|tensor| tensor_view_to_array_view(tensor))?
+    }
+
+    /// Get a tensor with specific dimensions by name
+    ///
+    /// # Example
+    /// ```rust
+    /// # use ndarray::Array2;
+    /// # use ndarray_safetensors::{SafeArrays, SafeArrayView};
+    /// # let array = Array2::<f32>::ones((2, 3));
+    /// # let mut safe_arrays = SafeArrays::new();
+    /// # safe_arrays.insert_ndarray("data", array.view()).unwrap();
+    /// # let bytes = safe_arrays.serialize().unwrap();
+    /// # let view = SafeArrayView::from_bytes(&bytes).unwrap();
+    /// let tensor: ndarray::ArrayView2<f32> = view.tensor("data").unwrap();
+    /// ```
+    pub fn tensor<T: STDtype, Dim: ndarray::Dimension>(
+        &self,
+        name: &str,
+    ) -> Result<ndarray::ArrayView<'a, T, Dim>> {
+        Ok(self
+            .tensors
+            .tensor(name)
+            .map(|tensor| tensor_view_to_array_view(tensor))?
+            .map(|array_view| array_view.into_dimensionality::<Dim>())??)
+    }
+
+    pub fn tensor_by_index<T: STDtype, Dim: ndarray::Dimension>(
+        &self,
+        index: usize,
+    ) -> Result<ndarray::ArrayView<'a, T, Dim>> {
+        self.tensors
+            .iter()
+            .nth(index)
+            .ok_or(SafeTensorError::TensorNotFound(format!(
+                "Index {} out of bounds",
+                index
+            )))
+            .map(|(_, tensor)| tensor_view_to_array_view(tensor))?
+            .map(|array_view| array_view.into_dimensionality::<Dim>())?
+            .map_err(SafeTensorError::NdarrayShapeError)
+    }
+
+    /// Get an iterator over tensor names
+    pub fn names(&self) -> std::vec::IntoIter<&str> {
+        self.tensors.names().into_iter()
+    }
+
+    /// Get the number of tensors
+    pub fn len(&self) -> usize {
+        self.tensors.len()
+    }
+
+    /// Check if there are no tensors
+    pub fn is_empty(&self) -> bool {
+        self.tensors.is_empty()
+    }
+}
+
+/// Trait for types that can be stored in SafeTensors
+///
+/// Implemented for: f32, f64, i8, i16, i32, i64, u8, u16, u32, u64, f16, bf16
+pub trait STDtype: bytemuck::Pod {
+    fn dtype() -> safetensors::tensor::Dtype;
+    fn size() -> usize {
+        (Self::dtype().bitsize() / 8).max(1)
+    }
+}
+
+macro_rules! impl_dtype {
+    ($($t:ty => $dtype:expr),* $(,)?) => {
+        $(
+            impl STDtype for $t {
+                fn dtype() -> safetensors::tensor::Dtype {
+                    $dtype
+                }
+            }
+        )*
+    };
+}
+
+use safetensors::tensor::Dtype;
+
+impl_dtype!(
+    // bool => Dtype::BOOL, // idk if ndarray::ArrayD<bool> is packed
+    f32 => Dtype::F32,
+    f64 => Dtype::F64,
+    i8 => Dtype::I8,
+    i16 => Dtype::I16,
+    i32 => Dtype::I32,
+    i64 => Dtype::I64,
+    u8 => Dtype::U8,
+    u16 => Dtype::U16,
+    u32 => Dtype::U32,
+    u64 => Dtype::U64,
+    half::f16 => Dtype::F16,
+    half::bf16 => Dtype::BF16,
+);
+
+fn tensor_view_to_array_view<'a, T: STDtype>(
+    tensor: safetensors::tensor::TensorView<'a>,
+) -> Result<ndarray::ArrayViewD<'a, T>> {
+    let shape = tensor.shape();
+    let dtype = tensor.dtype();
+    if T::dtype() != dtype {
+        return Err(SafeTensorError::InvalidTensorData(
+            core::any::type_name::<T>(),
+            dtype.to_string(),
+        ));
+    }
+
+    let data = tensor.data();
+    let data: &[T] = bytemuck::cast_slice(data);
+    let array = ndarray::ArrayViewD::from_shape(shape, data)?;
+    Ok(array)
+}
+
+/// Builder for creating SafeTensors data from ndarray tensors
+///
+/// # Example
+/// ```rust
+/// use ndarray::{Array1, Array2};
+/// use ndarray_safetensors::SafeArrays;
+///
+/// let mut safe_arrays = SafeArrays::new();
+///
+/// let array1 = Array1::<f32>::from_vec(vec![1.0, 2.0, 3.0]);
+/// let array2 = Array2::<i32>::zeros((2, 2));
+///
+/// safe_arrays.insert_ndarray("vector", array1.view()).unwrap();
+/// safe_arrays.insert_ndarray("matrix", array2.view()).unwrap();
+/// safe_arrays.insert_metadata("version", "1.0");
+///
+/// let bytes = safe_arrays.serialize().unwrap();
+/// ```
+#[derive(Debug, Clone, Default)]
+#[non_exhaustive]
+pub struct SafeArrays<'a> {
+    pub tensors: BTreeMap<String, SafeArray<'a>>,
+    pub metadata: Option<HashMap<String, String>>,
+}
+
+impl<'a, K: AsRef<str>> FromIterator<(K, SafeArray<'a>)> for SafeArrays<'a> {
+    fn from_iter<T: IntoIterator<Item = (K, SafeArray<'a>)>>(iter: T) -> Self {
+        let tensors = iter
+            .into_iter()
+            .map(|(k, v)| (k.as_ref().to_owned(), v))
+            .collect();
+        Self {
+            tensors,
+            metadata: None,
+        }
+    }
+}
+
+impl<'a, K: AsRef<str>, T: IntoIterator<Item = (K, SafeArray<'a>)>> From<T> for SafeArrays<'a> {
+    fn from(iter: T) -> Self {
+        let tensors = iter
+            .into_iter()
+            .map(|(k, v)| (k.as_ref().to_owned(), v))
+            .collect();
+        Self {
+            tensors,
+            metadata: None,
+        }
+    }
+}
+
+impl<'a> SafeArrays<'a> {
+    /// Create a SafeArrays from an iterator of (name, ndarray::ArrayView) pairs
+    /// ```rust
+    /// use ndarray::{Array2, Array3};
+    /// use ndarray_safetensors::{SafeArrays, SafeArray};
+    /// let array = Array2::<f32>::zeros((3, 4));
+    /// let safe_arrays = SafeArrays::from_ndarrays(vec![
+    ///     ("test_tensor", array.view()),
+    ///     ("test_tensor2", array.view()),
+    /// ]).unwrap();
+    /// ```
+
+    pub fn from_ndarrays<
+        K: AsRef<str>,
+        T: STDtype,
+        D: ndarray::Dimension + 'a,
+        I: IntoIterator<Item = (K, ndarray::ArrayView<'a, T, D>)>,
+    >(
+        iter: I,
+    ) -> Result<Self> {
+        let tensors = iter
+            .into_iter()
+            .map(|(k, v)| Ok((k.as_ref().to_owned(), SafeArray::from_ndarray(v)?)))
+            .collect::<Result<BTreeMap<String, SafeArray<'a>>>>()?;
+        Ok(Self {
+            tensors,
+            metadata: None,
+        })
+    }
+}
+
+// impl<'a, K: AsRef<str>, T: IntoIterator<Item = (K, SafeArray<'a>)>> From<T> for SafeArrays<'a> {
+//     fn from(iter: T) -> Self {
+//         let tensors = iter
+//             .into_iter()
+//             .map(|(k, v)| (k.as_ref().to_owned(), v))
+//             .collect();
+//         Self {
+//             tensors,
+//             metadata: None,
+//         }
+//     }
+// }
+
+impl<'a> SafeArrays<'a> {
+    /// Create a new empty SafeArrays builder
+    pub const fn new() -> Self {
+        Self {
+            tensors: BTreeMap::new(),
+            metadata: None,
+        }
+    }
+
+    /// Insert a SafeArray tensor with the given name
+    pub fn insert_tensor<'b: 'a>(&mut self, name: impl AsRef<str>, tensor: SafeArray<'b>) {
+        self.tensors.insert(name.as_ref().to_owned(), tensor);
+    }
+
+    /// Insert an ndarray tensor with the given name
+    ///
+    /// The array must be in standard layout and contiguous.
+    pub fn insert_ndarray<'b: 'a, T: STDtype, D: ndarray::Dimension + 'a>(
+        &mut self,
+        name: impl AsRef<str>,
+        array: ndarray::ArrayView<'b, T, D>,
+    ) -> Result<()> {
+        self.insert_tensor(name, SafeArray::from_ndarray(array)?);
+        Ok(())
+    }
+
+    /// Insert metadata key-value pair
+    pub fn insert_metadata(&mut self, key: impl AsRef<str>, value: impl AsRef<str>) {
+        self.metadata
+            .get_or_insert_default()
+            .insert(key.as_ref().to_owned(), value.as_ref().to_owned());
+    }
+
+    /// Serialize all tensors and metadata to bytes
+    pub fn serialize(self) -> Result<Vec<u8>> {
+        let out = safetensors::serialize(self.tensors, self.metadata)
+            .map_err(SafeTensorError::SafeTensor)?;
+        Ok(out)
+    }
+}
+
+/// A tensor that can be serialized to SafeTensors format
+#[derive(Debug, Clone)]
+pub struct SafeArray<'a> {
+    data: Cow<'a, [u8]>,
+    shape: Vec<usize>,
+    dtype: safetensors::tensor::Dtype,
+}
+
+impl View for SafeArray<'_> {
+    fn dtype(&self) -> safetensors::tensor::Dtype {
+        self.dtype
+    }
+
+    fn shape(&self) -> &[usize] {
+        &self.shape
+    }
+
+    fn data(&self) -> Cow<'_, [u8]> {
+        self.data.clone()
+    }
+
+    fn data_len(&self) -> usize {
+        self.data.len()
+    }
+}
+
+impl<'a> SafeArray<'a> {
+    fn from_ndarray<'b: 'a, T: STDtype, D: ndarray::Dimension + 'a>(
+        array: ndarray::ArrayView<'b, T, D>,
+    ) -> Result<Self> {
+        let shape = array.shape().to_vec();
+        let dtype = T::dtype();
+        if array.ndim() == 0 {
+            return Err(SafeTensorError::InvalidTensorData(
+                core::any::type_name::<T>(),
+                "Cannot insert a scalar tensor".to_string(),
+            ));
+        }
+
+        if !array.is_standard_layout() {
+            return Err(SafeTensorError::InvalidTensorData(
+                core::any::type_name::<T>(),
+                "ArrayView is not standard layout".to_string(),
+            ));
+        }
+        let data =
+            bytemuck::cast_slice(array.to_slice().ok_or(SafeTensorError::InvalidTensorData(
+                core::any::type_name::<T>(),
+                "ArrayView is not contiguous".to_string(),
+            ))?);
+        let safe_array = SafeArray {
+            data: Cow::Borrowed(data),
+            shape,
+            dtype,
+        };
+        Ok(safe_array)
+    }
+}
+
+#[test]
+fn test_safe_array_from_ndarray() {
+    use ndarray::Array2;
+
+    let array = Array2::<f32>::zeros((3, 4));
+    let safe_array = SafeArray::from_ndarray(array.view()).unwrap();
+    assert_eq!(safe_array.shape, vec![3, 4]);
+    assert_eq!(safe_array.dtype, safetensors::tensor::Dtype::F32);
+    assert_eq!(safe_array.data.len(), 3 * 4 * 4); // 3x4x4 bytes for f32
+}
+
+#[test]
+fn test_serialize_safe_arrays() {
+    use ndarray::{Array2, Array3};
+
+    let mut safe_arrays = SafeArrays::new();
+    let array = Array2::<f32>::zeros((3, 4));
+    let array2 = Array3::<u16>::zeros((8, 1, 9));
+    safe_arrays
+        .insert_ndarray("test_tensor", array.view())
+        .unwrap();
+    safe_arrays
+        .insert_ndarray("test_tensor2", array2.view())
+        .unwrap();
+    safe_arrays.insert_metadata("author", "example");
+
+    let serialized = safe_arrays.serialize().unwrap();
+    assert!(!serialized.is_empty());
+
+    // Deserialize to check if it works
+    let deserialized = SafeArraysView::from_bytes(&serialized).unwrap();
+    assert_eq!(deserialized.len(), 2);
+    assert_eq!(
+        deserialized
+            .tensor::<f32, ndarray::Ix2>("test_tensor")
+            .unwrap()
+            .shape(),
+        &[3, 4]
+    );
+    assert_eq!(
+        deserialized
+            .tensor::<u16, ndarray::Ix3>("test_tensor2")
+            .unwrap()
+            .shape(),
+        &[8, 1, 9]
+    );
+}

ndcv-bridge/Cargo.toml

@@ -0,0 +1,36 @@
+[package]
+name = "ndcv-bridge"
+version.workspace = true
+edition.workspace = true
+
+[dependencies]
+bounding-box.workspace = true
+nalgebra.workspace = true
+bytemuck.workspace = true
+error-stack.workspace = true
+fast_image_resize.workspace = true
+ndarray = { workspace = true, features = ["rayon"] }
+num.workspace = true
+opencv = { workspace = true, optional = true }
+rayon = "1.10.0"
+thiserror.workspace = true
+tracing = "0.1.41"
+wide = "0.7.32"
+img-parts.workspace = true
+
+[dev-dependencies]
+divan.workspace = true
+ndarray-npy.workspace = true
+
+[features]
+opencv = ["dep:opencv"]
+default = ["opencv"]
+
+
+[[bench]]
+name = "conversions"
+harness = false
+
+[[bench]]
+name = "gaussian"
+harness = false

ndcv-bridge/benches/conversions.rs

@@ -0,0 +1,75 @@
+use divan::black_box;
+use ndcv_bridge::*;
+
+// #[global_allocator]
+// static ALLOC: AllocProfiler = AllocProfiler::system();
+
+fn main() {
+    divan::main();
+}
+
+#[divan::bench]
+fn bench_3d_mat_to_ndarray_512() {
+    bench_mat_to_3d_ndarray(512);
+}
+
+#[divan::bench]
+fn bench_3d_mat_to_ndarray_1024() {
+    bench_mat_to_3d_ndarray(1024);
+}
+
+#[divan::bench]
+fn bench_3d_mat_to_ndarray_2k() {
+    bench_mat_to_3d_ndarray(2048);
+}
+
+#[divan::bench]
+fn bench_3d_mat_to_ndarray_4k() {
+    bench_mat_to_3d_ndarray(4096);
+}
+
+#[divan::bench]
+fn bench_3d_mat_to_ndarray_8k() {
+    bench_mat_to_3d_ndarray(8192);
+}
+
+#[divan::bench]
+fn bench_3d_mat_to_ndarray_8k_ref() {
+    bench_mat_to_3d_ndarray_ref(8192);
+}
+
+#[divan::bench]
+fn bench_2d_mat_to_ndarray_8k_ref() {
+    bench_mat_to_2d_ndarray(8192);
+}
+
+fn bench_mat_to_2d_ndarray(size: i32) -> ndarray::Array2<u8> {
+    let mat =
+        opencv::core::Mat::new_nd_with_default(&[size, size], opencv::core::CV_8UC1, (200).into())
+            .expect("failed");
+    let ndarray: ndarray::Array2<u8> = mat.as_ndarray().expect("failed").to_owned();
+    ndarray
+}
+
+fn bench_mat_to_3d_ndarray(size: i32) -> ndarray::Array3<u8> {
+    let mat = opencv::core::Mat::new_nd_with_default(
+        &[size, size],
+        opencv::core::CV_8UC3,
+        (200, 100, 10).into(),
+    )
+    .expect("failed");
+    // ndarray::Array3::<u8>::from_mat(black_box(mat)).expect("failed")
+    let ndarray: ndarray::Array3<u8> = mat.as_ndarray().expect("failed").to_owned();
+    ndarray
+}
+
+fn bench_mat_to_3d_ndarray_ref(size: i32) {
+    let mut mat = opencv::core::Mat::new_nd_with_default(
+        &[size, size],
+        opencv::core::CV_8UC3,
+        (200, 100, 10).into(),
+    )
+    .expect("failed");
+    let array: ndarray::ArrayView3<u8> = black_box(&mut mat).as_ndarray().expect("failed");
+    let _ = black_box(array);
+}

ndcv-bridge/benches/gaussian.rs

@@ -0,0 +1,265 @@
+use divan::black_box;
+use ndarray::*;
+use ndcv_bridge::*;
+
+// #[global_allocator]
+// static ALLOC: AllocProfiler = AllocProfiler::system();
+
+fn main() {
+    divan::main();
+}
+
+// Helper function to create test images with different patterns
+fn create_test_image(size: usize, pattern: &str) -> Array3<u8> {
+    let mut arr = Array3::<u8>::zeros((size, size, 3));
+    match pattern {
+        "edges" => {
+            // Create a pattern with sharp edges
+            arr.slice_mut(s![size / 4..3 * size / 4, size / 4..3 * size / 4, ..])
+                .fill(255);
+        }
+        "gradient" => {
+            // Create a gradual gradient
+            for i in 0..size {
+                let val = (i * 255 / size) as u8;
+                arr.slice_mut(s![i, .., ..]).fill(val);
+            }
+        }
+        "checkerboard" => {
+            // Create a checkerboard pattern
+            for i in 0..size {
+                for j in 0..size {
+                    if (i / 20 + j / 20) % 2 == 0 {
+                        arr[[i, j, 0]] = 255;
+                        arr[[i, j, 1]] = 255;
+                        arr[[i, j, 2]] = 255;
+                    }
+                }
+            }
+        }
+        _ => arr.fill(255), // Default to solid white
+    }
+    arr
+}
+
+#[divan::bench_group]
+mod sizes {
+    use super::*;
+    // Benchmark different image sizes
+    #[divan::bench(args = [512, 1024, 2048, 4096])]
+    fn bench_gaussian_sizes_u8(size: usize) {
+        let arr = Array3::<u8>::ones((size, size, 3));
+        let _out = black_box(
+            arr.gaussian_blur((3, 3), 1.0, 1.0, BorderType::BorderConstant)
+                .unwrap(),
+        );
+    }
+
+    #[divan::bench(args = [512, 1024, 2048, 4096])]
+    fn bench_gaussian_sizes_u8_inplace(size: usize) {
+        let mut arr = Array3::<u8>::ones((size, size, 3));
+        black_box(
+            arr.gaussian_blur_inplace((3, 3), 1.0, 1.0, BorderType::BorderConstant)
+                .unwrap(),
+        );
+    }
+
+    #[divan::bench(args = [512, 1024, 2048, 4096])]
+    fn bench_gaussian_sizes_f32(size: usize) {
+        let arr = Array3::<f32>::ones((size, size, 3));
+        let _out = black_box(
+            arr.gaussian_blur((3, 3), 1.0, 1.0, BorderType::BorderConstant)
+                .unwrap(),
+        );
+    }
+
+    #[divan::bench(args = [512, 1024, 2048, 4096])]
+    fn bench_gaussian_sizes_f32_inplace(size: usize) {
+        let mut arr = Array3::<f32>::ones((size, size, 3));
+        black_box(
+            arr.gaussian_blur_inplace((3, 3), 1.0, 1.0, BorderType::BorderConstant)
+                .unwrap(),
+        );
+    }
+}
+
+// Benchmark different kernel sizes
+#[divan::bench(args = [(3, 3), (5, 5), (7, 7), (9, 9), (11, 11)])]
+fn bench_gaussian_kernels(kernel_size: (u8, u8)) {
+    let mut arr = Array3::<u8>::ones((1000, 1000, 3));
+    arr.gaussian_blur_inplace(kernel_size, 1.0, 1.0, BorderType::BorderConstant)
+        .unwrap();
+}
+
+// Benchmark different sigma values
+#[divan::bench(args = [0.5, 1.0, 2.0, 5.0])]
+fn bench_gaussian_sigmas(sigma: f64) {
+    let mut arr = Array3::<u8>::ones((1000, 1000, 3));
+    arr.gaussian_blur_inplace((3, 3), sigma, sigma, BorderType::BorderConstant)
+        .unwrap();
+}
+
+// Benchmark different sigma_x and sigma_y combinations
+#[divan::bench(args = [(0.5, 2.0), (1.0, 1.0), (2.0, 0.5), (3.0, 1.0)])]
+fn bench_gaussian_asymmetric_sigmas(sigmas: (f64, f64)) {
+    let mut arr = Array3::<u8>::ones((1000, 1000, 3));
+    arr.gaussian_blur_inplace((3, 3), sigmas.0, sigmas.1, BorderType::BorderConstant)
+        .unwrap();
+}
+
+// Benchmark different border types
+#[divan::bench]
+fn bench_gaussian_border_types() -> Vec<()> {
+    let border_types = [
+        BorderType::BorderConstant,
+        BorderType::BorderReplicate,
+        BorderType::BorderReflect,
+        BorderType::BorderReflect101,
+    ];
+
+    let mut arr = Array3::<u8>::ones((1000, 1000, 3));
+    border_types
+        .iter()
+        .map(|border_type| {
+            arr.gaussian_blur_inplace((3, 3), 1.0, 1.0, *border_type)
+                .unwrap();
+        })
+        .collect()
+}
+
+// Benchmark different image patterns
+#[divan::bench]
+fn bench_gaussian_patterns() {
+    let patterns = ["edges", "gradient", "checkerboard", "solid"];
+
+    patterns.iter().for_each(|&pattern| {
+        let mut arr = create_test_image(1000, pattern);
+        arr.gaussian_blur_inplace((3, 3), 1.0, 1.0, BorderType::BorderConstant)
+            .unwrap();
+    })
+}
+
+#[divan::bench_group]
+mod allocation {
+    use super::*;
+    #[divan::bench]
+    fn bench_gaussian_allocation_inplace() {
+        let mut arr = Array3::<f32>::ones((3840, 2160, 3));
+
+        black_box(
+            arr.gaussian_blur_inplace((3, 3), 1.0, 1.0, BorderType::BorderConstant)
+                .unwrap(),
+        );
+    }
+
+    #[divan::bench]
+    fn bench_gaussian_allocation_allocate() {
+        let arr = Array3::<f32>::ones((3840, 2160, 3));
+
+        let _out = black_box(
+            arr.gaussian_blur((3, 3), 1.0, 1.0, BorderType::BorderConstant)
+                .unwrap(),
+        );
+    }
+}
+
+#[divan::bench_group]
+mod realistic {
+    use super::*;
+    #[divan::bench]
+    fn small_800_600_3x3() {
+        let small_blur = Array3::<u8>::ones((800, 600, 3));
+        let _blurred = black_box(
+            small_blur
+                .gaussian_blur((3, 3), 0.5, 0.5, BorderType::BorderConstant)
+                .unwrap(),
+        );
+    }
+    #[divan::bench]
+    fn small_800_600_3x3_inplace() {
+        let mut small_blur = Array3::<u8>::ones((800, 600, 3));
+        small_blur
+            .gaussian_blur_inplace((3, 3), 0.5, 0.5, BorderType::BorderConstant)
+            .unwrap();
+    }
+    #[divan::bench]
+    fn medium_1920x1080_5x5() {
+        let mut medium_blur = Array3::<u8>::ones((1920, 1080, 3));
+        let _blurred = black_box(
+            medium_blur
+                .gaussian_blur_inplace((5, 5), 2.0, 2.0, BorderType::BorderConstant)
+                .unwrap(),
+        );
+    }
+    #[divan::bench]
+    fn medium_1920x1080_5x5_inplace() {
+        let mut medium_blur = Array3::<u8>::ones((1920, 1080, 3));
+        medium_blur
+            .gaussian_blur_inplace((5, 5), 2.0, 2.0, BorderType::BorderConstant)
+            .unwrap();
+    }
+    #[divan::bench]
+    fn large_3840x2160_9x9() {
+        let large_blur = Array3::<u8>::ones((3840, 2160, 3));
+        let _blurred = black_box(
+            large_blur
+                .gaussian_blur((9, 9), 5.0, 5.0, BorderType::BorderConstant)
+                .unwrap(),
+        );
+    }
+    #[divan::bench]
+    fn large_3840x2160_9x9_inplace() {
+        let mut large_blur = Array3::<u8>::ones((3840, 2160, 3));
+        large_blur
+            .gaussian_blur_inplace((9, 9), 5.0, 5.0, BorderType::BorderConstant)
+            .unwrap();
+    }
+    #[divan::bench]
+    fn small_800_600_3x3_f32() {
+        let small_blur = Array3::<f32>::ones((800, 600, 3));
+        let _blurred = black_box(
+            small_blur
+                .gaussian_blur((3, 3), 0.5, 0.5, BorderType::BorderConstant)
+                .unwrap(),
+        );
+    }
+    #[divan::bench]
+    fn small_800_600_3x3_inplace_f32() {
+        let mut small_blur = Array3::<f32>::ones((800, 600, 3));
+        small_blur
+            .gaussian_blur_inplace((3, 3), 0.5, 0.5, BorderType::BorderConstant)
+            .unwrap();
+    }
+    #[divan::bench]
+    fn medium_1920x1080_5x5_f32() {
+        let mut medium_blur = Array3::<f32>::ones((1920, 1080, 3));
+        let _blurred = black_box(
+            medium_blur
+                .gaussian_blur_inplace((5, 5), 2.0, 2.0, BorderType::BorderConstant)
+                .unwrap(),
+        );
+    }
+    #[divan::bench]
+    fn medium_1920x1080_5x5_inplace_f32() {
+        let mut medium_blur = Array3::<f32>::ones((1920, 1080, 3));
+        medium_blur
+            .gaussian_blur_inplace((5, 5), 2.0, 2.0, BorderType::BorderConstant)
+            .unwrap();
+    }
+    #[divan::bench]
+    fn large_3840x2160_9x9_f32() {
+        let large_blur = Array3::<f32>::ones((3840, 2160, 3));
+        let _blurred = black_box(
+            large_blur
+                .gaussian_blur((9, 9), 5.0, 5.0, BorderType::BorderConstant)
+                .unwrap(),
+        );
+    }
+    #[divan::bench]
+    fn large_3840x2160_9x9_inplace_f32() {
+        let mut large_blur = Array3::<f32>::ones((3840, 2160, 3));
+        large_blur
+            .gaussian_blur_inplace((9, 9), 5.0, 5.0, BorderType::BorderConstant)
+            .unwrap();
+    }
+}

ndcv-bridge/src/blend.rs

@@ -0,0 +1,180 @@
+use crate::prelude_::*;
+use ndarray::*;
+
+type Result<T, E = Report<NdCvError>> = std::result::Result<T, E>;
+
+mod seal {
+    pub trait Sealed {}
+    impl<T: ndarray::Data<Elem = f32>> Sealed for ndarray::ArrayBase<T, ndarray::Ix3> {}
+}
+pub trait NdBlend<T, D: ndarray::Dimension>: seal::Sealed {
+    fn blend(
+        &self,
+        mask: ndarray::ArrayView<T, D::Smaller>,
+        other: ndarray::ArrayView<T, D>,
+        alpha: T,
+    ) -> Result<ndarray::Array<T, D>>;
+    fn blend_inplace(
+        &mut self,
+        mask: ndarray::ArrayView<T, D::Smaller>,
+        other: ndarray::ArrayView<T, D>,
+        alpha: T,
+    ) -> Result<()>;
+}
+
+impl<S> NdBlend<f32, Ix3> for ndarray::ArrayBase<S, Ix3>
+where
+    S: ndarray::DataMut<Elem = f32>,
+{
+    fn blend(
+        &self,
+        mask: ndarray::ArrayView<f32, Ix2>,
+        other: ndarray::ArrayView<f32, Ix3>,
+        alpha: f32,
+    ) -> Result<ndarray::Array<f32, Ix3>> {
+        if self.shape() != other.shape() {
+            return Err(NdCvError)
+                .attach_printable("Shapes of image and other imagge do not match");
+        }
+        if self.shape()[0] != mask.shape()[0] || self.shape()[1] != mask.shape()[1] {
+            return Err(NdCvError).attach_printable("Shapes of image and mask do not match");
+        }
+
+        let mut output = ndarray::Array3::zeros(self.dim());
+        let (_height, _width, channels) = self.dim();
+
+        Zip::from(output.lanes_mut(Axis(2)))
+            .and(self.lanes(Axis(2)))
+            .and(other.lanes(Axis(2)))
+            .and(mask)
+            .par_for_each(|mut out, this, other, mask| {
+                let this = wide::f32x4::from(this.as_slice().expect("Invalid self array"));
+                let other = wide::f32x4::from(other.as_slice().expect("Invalid other array"));
+                let mask = wide::f32x4::splat(mask * alpha);
+                let o = this * (1.0 - mask) + other * mask;
+                out.as_slice_mut()
+                    .expect("Failed to get mutable slice")
+                    .copy_from_slice(&o.as_array_ref()[..channels]);
+            });
+
+        Ok(output)
+    }
+
+    fn blend_inplace(
+        &mut self,
+        mask: ndarray::ArrayView<f32, <Ix3 as Dimension>::Smaller>,
+        other: ndarray::ArrayView<f32, Ix3>,
+        alpha: f32,
+    ) -> Result<()> {
+        if self.shape() != other.shape() {
+            return Err(NdCvError)
+                .attach_printable("Shapes of image and other imagge do not match");
+        }
+        if self.shape()[0] != mask.shape()[0] || self.shape()[1] != mask.shape()[1] {
+            return Err(NdCvError).attach_printable("Shapes of image and mask do not match");
+        }
+
+        let (_height, _width, channels) = self.dim();
+
+        // Zip::from(self.lanes_mut(Axis(2)))
+        //     .and(other.lanes(Axis(2)))
+        //     .and(mask)
+        //     .par_for_each(|mut this, other, mask| {
+        //         let this_wide = wide::f32x4::from(this.as_slice().expect("Invalid self array"));
+        //         let other = wide::f32x4::from(other.as_slice().expect("Invalid other array"));
+        //         let mask = wide::f32x4::splat(mask * alpha);
+        //         let o = this_wide * (1.0 - mask) + other * mask;
+        //         this.as_slice_mut()
+        //             .expect("Failed to get mutable slice")
+        //             .copy_from_slice(&o.as_array_ref()[..channels]);
+        //     });
+        let this = self
+            .as_slice_mut()
+            .ok_or(NdCvError)
+            .attach_printable("Failed to get source image as a continuous slice")?;
+        let other = other
+            .as_slice()
+            .ok_or(NdCvError)
+            .attach_printable("Failed to get other image as a continuous slice")?;
+        let mask = mask
+            .as_slice()
+            .ok_or(NdCvError)
+            .attach_printable("Failed to get mask as a continuous slice")?;
+
+        use rayon::prelude::*;
+        this.par_chunks_exact_mut(channels)
+            .zip(other.par_chunks_exact(channels))
+            .zip(mask)
+            .for_each(|((this, other), mask)| {
+                let this_wide = wide::f32x4::from(&*this);
+                let other = wide::f32x4::from(other);
+                let mask = wide::f32x4::splat(mask * alpha);
+                this.copy_from_slice(
+                    &(this_wide * (1.0 - mask) + other * mask).as_array_ref()[..channels],
+                );
+            });
+
+        // for h in 0.._height {
+        //     for w in 0.._width {
+        //         let mask_index = h * _width + w;
+        //         let mask = mask[mask_index];
+        //         let mask = wide::f32x4::splat(mask * alpha);
+        //         let this = &mut this[mask_index * channels..(mask_index + 1) * channels];
+        //         let other = &other[mask_index * channels..(mask_index + 1) * channels];
+        //         let this_wide = wide::f32x4::from(&*this);
+        //         let other = wide::f32x4::from(other);
+        //         let o = this_wide * (1.0 - mask) + other * mask;
+        //         this.copy_from_slice(&o.as_array_ref()[..channels]);
+        //     }
+        // }
+        Ok(())
+    }
+}
+
+#[test]
+pub fn test_blend() {
+    let img = Array3::<f32>::from_shape_fn((10, 10, 3), |(i, j, k)| match (i, j, k) {
+        (0..=3, _, 0) => 1f32, // red
+        (4..=6, _, 1) => 1f32, // green
+        (7..=9, _, 2) => 1f32, // blue
+        _ => 0f32,
+    });
+    let other = img.clone().permuted_axes([1, 0, 2]).to_owned();
+    let mask = Array2::<f32>::from_shape_fn((10, 10), |(_, j)| if j > 5 { 1f32 } else { 0f32 });
+    // let other = Array3::<f32>::zeros((10, 10, 3));
+    let out = img.blend(mask.view(), other.view(), 1f32).unwrap();
+    let out_u8 = out.mapv(|v| (v * 255f32) as u8);
+    let expected = Array3::<u8>::from_shape_fn((10, 10, 3), |(i, j, k)| {
+        match (i, j, k) {
+            (0..=3, 0..=5, 0) => u8::MAX,                  // red
+            (4..=6, 0..=5, 1) | (_, 6, 1) => u8::MAX,      // green
+            (7..=9, 0..=5, 2) | (_, 7..=10, 2) => u8::MAX, // blue
+            _ => u8::MIN,
+        }
+    });
+    assert_eq!(out_u8, expected);
+}
+
+// #[test]
+// pub fn test_blend_inplace() {
+//     let mut img = Array3::<f32>::from_shape_fn((10, 10, 3), |(i, j, k)| match (i, j, k) {
+//         (0..=3, _, 0) => 1f32, // red
+//         (4..=6, _, 1) => 1f32, // green
+//         (7..=9, _, 2) => 1f32, // blue
+//         _ => 0f32,
+//     });
+//     let other = img.clone().permuted_axes([1, 0, 2]);
+//     let mask = Array2::<f32>::from_shape_fn((10, 10), |(_, j)| if j > 5 { 1f32 } else { 0f32 });
+//     // let other = Array3::<f32>::zeros((10, 10, 3));
+//     img.blend_inplace(mask.view(), other.view(), 1f32).unwrap();
+//     let out_u8 = img.mapv(|v| (v * 255f32) as u8);
+//     let expected = Array3::<u8>::from_shape_fn((10, 10, 3), |(i, j, k)| {
+//         match (i, j, k) {
+//             (0..=3, 0..=5, 0) => u8::MAX,                  // red
+//             (4..=6, 0..=5, 1) | (_, 6, 1) => u8::MAX,      // green
+//             (7..=9, 0..=5, 2) | (_, 7..=10, 2) => u8::MAX, // blue
+//             _ => u8::MIN,
+//         }
+//     });
+//     assert_eq!(out_u8, expected);
+// }

ndcv-bridge/src/bounding_rect.rs

@@ -0,0 +1,48 @@
+//! Calculates the up-right bounding rectangle of a point set or non-zero pixels of gray-scale image.
+//! The function calculates and returns the minimal up-right bounding rectangle for the specified point set or non-zero pixels of gray-scale image.
+use crate::{NdAsImage, prelude_::*};
+pub trait BoundingRect: seal::SealedInternal {
+    fn bounding_rect(&self) -> Result<bounding_box::Aabb2<i32>, NdCvError>;
+}
+
+mod seal {
+    pub trait SealedInternal {}
+    impl<T, S: ndarray::Data<Elem = T>> SealedInternal for ndarray::ArrayBase<S, ndarray::Ix2> {}
+}
+
+impl<S: ndarray::Data<Elem = u8>> BoundingRect for ndarray::ArrayBase<S, ndarray::Ix2> {
+    fn bounding_rect(&self) -> Result<bounding_box::Aabb2<i32>, NdCvError> {
+        let mat = self.as_image_mat()?;
+        let rect = opencv::imgproc::bounding_rect(mat.as_ref()).change_context(NdCvError)?;
+        Ok(bounding_box::Aabb2::from_xywh(
+            rect.x,
+            rect.y,
+            rect.width,
+            rect.height,
+        ))
+    }
+}
+
+#[test]
+fn test_bounding_rect_empty() {
+    let arr = ndarray::Array2::<u8>::zeros((10, 10));
+    let rect = arr.bounding_rect().unwrap();
+    assert_eq!(rect, bounding_box::Aabb2::from_xywh(0, 0, 0, 0));
+}
+
+#[test]
+fn test_bounding_rect_valued() {
+    let mut arr = ndarray::Array2::<u8>::zeros((10, 10));
+    crate::NdRoiMut::roi_mut(&mut arr, bounding_box::Aabb2::from_xywh(1, 1, 3, 3)).fill(1);
+    let rect = arr.bounding_rect().unwrap();
+    assert_eq!(rect, bounding_box::Aabb2::from_xywh(1, 1, 3, 3));
+}
+
+#[test]
+fn test_bounding_rect_complex() {
+    let mut arr = ndarray::Array2::<u8>::zeros((10, 10));
+    crate::NdRoiMut::roi_mut(&mut arr, bounding_box::Aabb2::from_xywh(1, 3, 3, 3)).fill(1);
+    crate::NdRoiMut::roi_mut(&mut arr, bounding_box::Aabb2::from_xywh(2, 3, 3, 5)).fill(5);
+    let rect = arr.bounding_rect().unwrap();
+    assert_eq!(rect, bounding_box::Aabb2::from_xywh(1, 3, 4, 5));
+}

ndcv-bridge/src/codec.rs

@@ -0,0 +1,4 @@
+pub mod codecs;
+pub mod decode;
+pub mod encode;
+pub mod error;

ndcv-bridge/src/codec/codecs.rs

@@ -0,0 +1,218 @@
+use super::decode::Decoder;
+use super::encode::Encoder;
+use crate::NdCvError;
+use crate::conversions::matref::MatRef;
+use error_stack::*;
+use img_parts::{
+    Bytes,
+    jpeg::{Jpeg, markers},
+};
+use opencv::{
+    core::{Mat, Vector, VectorToVec},
+    imgcodecs::{ImreadModes, ImwriteFlags, imdecode, imencode},
+};
+
+#[derive(Debug)]
+pub enum CvEncoder {
+    Jpeg(CvJpegEncFlags),
+    Tiff(CvTiffEncFlags),
+}
+
+pub enum EncKind {
+    Jpeg,
+    Tiff,
+}
+
+impl CvEncoder {
+    fn kind(&self) -> EncKind {
+        match self {
+            Self::Jpeg(_) => EncKind::Jpeg,
+            Self::Tiff(_) => EncKind::Tiff,
+        }
+    }
+
+    fn extension(&self) -> &'static str {
+        match self {
+            Self::Jpeg(_) => ".jpg",
+            Self::Tiff(_) => ".tiff",
+        }
+    }
+
+    fn to_cv_param_list(&self) -> Vector<i32> {
+        match self {
+            Self::Jpeg(flags) => flags.to_cv_param_list(),
+            Self::Tiff(flags) => flags.to_cv_param_list(),
+        }
+    }
+}
+
+#[derive(Default, Debug)]
+pub struct CvJpegEncFlags {
+    quality: Option<usize>,
+    progressive: Option<bool>,
+    optimize: Option<bool>,
+    remove_app0: Option<bool>,
+}
+
+#[derive(Default, Debug)]
+pub struct CvTiffEncFlags {
+    compression: Option<i32>,
+}
+
+impl CvTiffEncFlags {
+    pub fn new() -> Self {
+        Self::default().with_compression(1)
+    }
+
+    pub fn with_compression(mut self, compression: i32) -> Self {
+        self.compression = Some(compression);
+        self
+    }
+
+    fn to_cv_param_list(&self) -> Vector<i32> {
+        let iter = [(
+            ImwriteFlags::IMWRITE_TIFF_COMPRESSION as i32,
+            self.compression.map(|i| i as i32),
+        )]
+        .into_iter()
+        .filter_map(|(flag, opt)| opt.map(|o| [flag, o]))
+        .flatten();
+
+        Vector::from_iter(iter)
+    }
+}
+
+impl CvJpegEncFlags {
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    pub fn with_quality(mut self, quality: usize) -> Self {
+        self.quality = Some(quality);
+        self
+    }
+
+    pub fn remove_app0_marker(mut self, val: bool) -> Self {
+        self.remove_app0 = Some(val);
+        self
+    }
+
+    fn to_cv_param_list(&self) -> Vector<i32> {
+        let iter = [
+            (
+                ImwriteFlags::IMWRITE_JPEG_QUALITY as i32,
+                self.quality.map(|i| i as i32),
+            ),
+            (
+                ImwriteFlags::IMWRITE_JPEG_PROGRESSIVE as i32,
+                self.progressive.map(|i| i as i32),
+            ),
+            (
+                ImwriteFlags::IMWRITE_JPEG_OPTIMIZE as i32,
+                self.optimize.map(|i| i as i32),
+            ),
+        ]
+        .into_iter()
+        .filter_map(|(flag, opt)| opt.map(|o| [flag, o]))
+        .flatten();
+
+        Vector::from_iter(iter)
+    }
+}
+
+impl Encoder for CvEncoder {
+    type Input<'a>
+        = MatRef<'a>
+    where
+        Self: 'a;
+
+    fn encode(&self, input: Self::Input<'_>) -> Result<Vec<u8>, NdCvError> {
+        let mut buf = Vector::default();
+
+        let params = self.to_cv_param_list();
+
+        imencode(self.extension(), &input.as_ref(), &mut buf, &params).change_context(NdCvError)?;
+
+        match self.kind() {
+            EncKind::Jpeg => {
+                let bytes = Bytes::from(buf.to_vec());
+                let mut jpg = Jpeg::from_bytes(bytes).change_context(NdCvError)?;
+                jpg.remove_segments_by_marker(markers::APP0);
+                let bytes = jpg.encoder().bytes();
+                Ok(bytes.to_vec())
+            }
+            EncKind::Tiff => Ok(buf.to_vec()),
+        }
+    }
+}
+
+pub enum CvDecoder {
+    Jpeg(CvJpegDecFlags),
+}
+
+impl CvDecoder {
+    fn to_cv_decode_flag(&self) -> i32 {
+        match self {
+            Self::Jpeg(flags) => flags.to_cv_decode_flag(),
+        }
+    }
+}
+
+#[derive(Default)]
+pub enum ColorMode {
+    #[default]
+    Color,
+    GrayScale,
+}
+
+impl ColorMode {
+    fn to_cv_decode_flag(&self) -> i32 {
+        match self {
+            Self::Color => ImreadModes::IMREAD_ANYCOLOR as i32,
+            Self::GrayScale => ImreadModes::IMREAD_GRAYSCALE as i32,
+        }
+    }
+}
+
+#[derive(Default)]
+pub struct CvJpegDecFlags {
+    color_mode: ColorMode,
+    ignore_orientation: bool,
+}
+
+impl CvJpegDecFlags {
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    pub fn with_color_mode(mut self, color_mode: ColorMode) -> Self {
+        self.color_mode = color_mode;
+        self
+    }
+
+    pub fn with_ignore_orientation(mut self, ignore_orientation: bool) -> Self {
+        self.ignore_orientation = ignore_orientation;
+        self
+    }
+
+    fn to_cv_decode_flag(&self) -> i32 {
+        let flag = self.color_mode.to_cv_decode_flag();
+
+        if self.ignore_orientation {
+            flag | ImreadModes::IMREAD_IGNORE_ORIENTATION as i32
+        } else {
+            flag
+        }
+    }
+}
+
+impl Decoder for CvDecoder {
+    type Output = Mat;
+
+    fn decode(&self, input: impl AsRef<[u8]>) -> Result<Self::Output, NdCvError> {
+        let flag = self.to_cv_decode_flag();
+        let out = imdecode(&Vector::from_slice(input.as_ref()), flag).change_context(NdCvError)?;
+
+        Ok(out)
+    }
+}

ndcv-bridge/src/codec/decode.rs

@@ -0,0 +1,61 @@
+#![deny(warnings)]
+
+use super::codecs::CvDecoder;
+use super::error::ErrorReason;
+use crate::NdCvError;
+use crate::{NdAsImage, conversions::NdCvConversion};
+use error_stack::*;
+use ndarray::Array;
+use std::path::Path;
+
+pub trait Decodable<D: Decoder>: Sized {
+    fn decode(buf: impl AsRef<[u8]>, decoder: &D) -> Result<Self, NdCvError> {
+        let output = decoder.decode(buf)?;
+        Self::transform(output)
+    }
+
+    fn read(&self, path: impl AsRef<Path>, decoder: &D) -> Result<Self, NdCvError> {
+        let buf = std::fs::read(path)
+            .map_err(|e| match e.kind() {
+                std::io::ErrorKind::NotFound => {
+                    Report::new(e).attach_printable(ErrorReason::ImageWriteFileNotFound)
+                }
+                std::io::ErrorKind::PermissionDenied => {
+                    Report::new(e).attach_printable(ErrorReason::ImageWritePermissionDenied)
+                }
+                std::io::ErrorKind::OutOfMemory => {
+                    Report::new(e).attach_printable(ErrorReason::OutOfMemory)
+                }
+                std::io::ErrorKind::StorageFull => {
+                    Report::new(e).attach_printable(ErrorReason::OutOfStorage)
+                }
+                _ => Report::new(e).attach_printable(ErrorReason::ImageWriteOtherError),
+            })
+            .change_context(NdCvError)?;
+        Self::decode(buf, decoder)
+    }
+
+    fn transform(input: D::Output) -> Result<Self, NdCvError>;
+}
+
+pub trait Decoder {
+    type Output: Sized;
+    fn decode(&self, buf: impl AsRef<[u8]>) -> Result<Self::Output, NdCvError>;
+}
+
+impl<T: bytemuck::Pod + Copy, D: ndarray::Dimension> Decodable<CvDecoder> for Array<T, D>
+where
+    Self: NdAsImage<T, D>,
+{
+    fn transform(input: <CvDecoder as Decoder>::Output) -> Result<Self, NdCvError> {
+        Self::from_mat(input)
+    }
+}
+
+#[test]
+fn decode_image() {
+    use crate::codec::codecs::*;
+    let img = std::fs::read("/Users/fs0c131y/Projects/face-detector/assets/selfie.jpg").unwrap();
+    let decoder = CvDecoder::Jpeg(CvJpegDecFlags::new().with_ignore_orientation(true));
+    let _out = ndarray::Array3::<u8>::decode(img, &decoder).unwrap();
+}

ndcv-bridge/src/codec/encode.rs

@@ -0,0 +1,56 @@
+use super::codecs::CvEncoder;
+use super::error::ErrorReason;
+use crate::conversions::NdAsImage;
+use crate::NdCvError;
+use error_stack::*;
+use ndarray::ArrayBase;
+use std::path::Path;
+
+pub trait Encodable<E: Encoder> {
+    fn encode(&self, encoder: &E) -> Result<Vec<u8>, NdCvError> {
+        let input = self.transform()?;
+        encoder.encode(input)
+    }
+
+    fn write(&self, path: impl AsRef<Path>, encoder: &E) -> Result<(), NdCvError> {
+        let buf = self.encode(encoder)?;
+
+        std::fs::write(path, buf)
+            .map_err(|e| match e.kind() {
+                std::io::ErrorKind::NotFound => {
+                    Report::new(e).attach_printable(ErrorReason::ImageWriteFileNotFound)
+                }
+                std::io::ErrorKind::PermissionDenied => {
+                    Report::new(e).attach_printable(ErrorReason::ImageWritePermissionDenied)
+                }
+                std::io::ErrorKind::OutOfMemory => {
+                    Report::new(e).attach_printable(ErrorReason::OutOfMemory)
+                }
+                std::io::ErrorKind::StorageFull => {
+                    Report::new(e).attach_printable(ErrorReason::OutOfStorage)
+                }
+                _ => Report::new(e).attach_printable(ErrorReason::ImageWriteOtherError),
+            })
+            .change_context(NdCvError)
+    }
+
+    fn transform(&self) -> Result<<E as Encoder>::Input<'_>, NdCvError>;
+}
+
+pub trait Encoder {
+    type Input<'a>
+    where
+        Self: 'a;
+
+    fn encode(&self, input: Self::Input<'_>) -> Result<Vec<u8>, NdCvError>;
+}
+
+impl<T: bytemuck::Pod + Copy, S: ndarray::Data<Elem = T>, D: ndarray::Dimension>
+    Encodable<CvEncoder> for ArrayBase<S, D>
+where
+    Self: NdAsImage<T, D>,
+{
+    fn transform(&self) -> Result<<CvEncoder as Encoder>::Input<'_>, NdCvError> {
+        self.as_image_mat()
+    }
+}

ndcv-bridge/src/codec/error.rs

@@ -0,0 +1,19 @@
+#[derive(Debug)]
+pub enum ErrorReason {
+    ImageReadFileNotFound,
+    ImageReadPermissionDenied,
+    ImageReadOtherError,
+
+    ImageWriteFileNotFound,
+    ImageWritePermissionDenied,
+    ImageWriteOtherError,
+
+    OutOfMemory,
+    OutOfStorage,
+}
+
+impl std::fmt::Display for ErrorReason {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{:?}", self)
+    }
+}

ndcv-bridge/src/color_space.rs

@@ -0,0 +1,88 @@
+//! Colorspace conversion functions
+//! ## Example
+//! ```rust
+//! let arr = Array3::<u8>::ones((100, 100, 3));
+//! let out: Array3<u8> = arr.cvt::<Rgba<u8>, Rgb<u8>>()
+//! ```
+use crate::prelude_::*;
+use ndarray::*;
+
+pub trait ColorSpace {
+    type Elem: seal::Sealed;
+    type Dim: ndarray::Dimension;
+    const CHANNELS: usize;
+}
+
+mod seal {
+    pub trait Sealed: bytemuck::Pod {}
+    // impl<T> Sealed for T {}
+    impl Sealed for u8 {} // 0 to 255
+    impl Sealed for u16 {} // 0 to 65535
+    impl Sealed for f32 {} // 0 to 1
+}
+
+macro_rules! define_color_space {
+    ($name:ident, $channels:expr, $depth:ty) => {
+        pub struct $name<T> {
+            __phantom: core::marker::PhantomData<T>,
+        }
+        impl<T: seal::Sealed> ColorSpace for $name<T> {
+            type Elem = T;
+            type Dim = $depth;
+            const CHANNELS: usize = $channels;
+        }
+    };
+}
+
+define_color_space!(Rgb, 3, Ix3);
+define_color_space!(Bgr, 3, Ix3);
+define_color_space!(Rgba, 4, Ix3);
+
+pub trait NdArray<T, D: ndarray::Dimension> {}
+impl<T, D: ndarray::Dimension, S: ndarray::Data<Elem = T>> NdArray<S, D> for ArrayBase<S, D> {}
+
+pub trait ConvertColor<T, U>
+where
+    T: ColorSpace,
+    U: ColorSpace,
+    Self: NdArray<T::Elem, T::Dim>,
+{
+    type Output: NdArray<U::Elem, U::Dim>;
+    fn cvt(&self) -> Self::Output;
+}
+
+// impl<T: seal::Sealed, S: ndarray::Data<Elem = T>> ConvertColor<Rgb<T>, Bgr<T>> for ArrayBase<S, Ix3>
+// where
+//     Self: NdArray<T, Ix3>,
+// {
+//     type Output = ArrayView3<'a, T>;
+//     fn cvt(&self) -> CowArray<T, Ix3> {
+//         self.view().permuted_axes([2, 1, 0]).into()
+//     }
+// }
+//
+// impl<T: seal::Sealed, S: ndarray::Data<Elem = T>> ConvertColor<Bgr<T>, Rgb<T>> for ArrayBase<S, Ix3>
+// where
+//     Self: NdArray<T, Ix3>,
+// {
+//     type Output = ArrayView3<'a, T>;
+//     fn cvt(&self) -> CowArray<T, Ix3> {
+//         self.view().permuted_axes([2, 1, 0]).into()
+//     }
+// }
+
+
+// impl<T: seal::Sealed + num::One + num::Zero, S: ndarray::Data<Elem = T>>
+//     ConvertColor<Rgb<T>, Rgba<T>> for ArrayBase<S, Ix3>
+// {
+//     fn cvt(&self) -> CowArray<T, Ix3> {
+//         let mut out = Array3::<T>::zeros((self.height(), self.width(), 4));
+//         // Zip::from(&mut out).and(self).for_each(|out, &in_| {
+//         //     out[0] = in_[0];
+//         //     out[1] = in_[1];
+//         //     out[2] = in_[2];
+//         //     out[3] = T::one();
+//         // });
+//         out.into()
+//     }
+// }

ndcv-bridge/src/connected_components.rs

@@ -0,0 +1,113 @@
+use crate::{NdAsImage, NdAsImageMut, conversions::MatAsNd, prelude_::*};
+
+pub(crate) mod seal {
+    pub trait ConnectedComponentOutput: Sized + Copy + bytemuck::Pod + num::Zero {
+        fn as_cv_type() -> i32 {
+            crate::type_depth::<Self>()
+        }
+    }
+    impl ConnectedComponentOutput for i32 {}
+    impl ConnectedComponentOutput for u16 {}
+}
+
+pub trait NdCvConnectedComponents<T> {
+    fn connected_components<O: seal::ConnectedComponentOutput>(
+        &self,
+        connectivity: Connectivity,
+    ) -> Result<ndarray::Array2<O>, NdCvError>;
+    fn connected_components_with_stats<O: seal::ConnectedComponentOutput>(
+        &self,
+        connectivity: Connectivity,
+    ) -> Result<ConnectedComponentStats<O>, NdCvError>;
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
+pub enum Connectivity {
+    Four = 4,
+    #[default]
+    Eight = 8,
+}
+
+#[derive(Debug, Clone)]
+pub struct ConnectedComponentStats<O: seal::ConnectedComponentOutput> {
+    pub num_labels: i32,
+    pub labels: ndarray::Array2<O>,
+    pub stats: ndarray::Array2<i32>,
+    pub centroids: ndarray::Array2<f64>,
+}
+
+// use crate::conversions::NdCvConversionRef;
+impl<T: bytemuck::Pod, S: ndarray::Data<Elem = T>> NdCvConnectedComponents<T>
+    for ndarray::ArrayBase<S, ndarray::Ix2>
+where
+    ndarray::Array2<T>: NdAsImage<T, ndarray::Ix2>,
+{
+    fn connected_components<O: seal::ConnectedComponentOutput>(
+        &self,
+        connectivity: Connectivity,
+    ) -> Result<ndarray::Array2<O>, NdCvError> {
+        let mat = self.as_image_mat()?;
+        let mut labels = ndarray::Array2::<O>::zeros(self.dim());
+        let mut cv_labels = labels.as_image_mat_mut()?;
+        opencv::imgproc::connected_components(
+            mat.as_ref(),
+            cv_labels.as_mut(),
+            connectivity as i32,
+            O::as_cv_type(),
+        )
+        .change_context(NdCvError)?;
+        Ok(labels)
+    }
+
+    fn connected_components_with_stats<O: seal::ConnectedComponentOutput>(
+        &self,
+        connectivity: Connectivity,
+    ) -> Result<ConnectedComponentStats<O>, NdCvError> {
+        let mut labels = ndarray::Array2::<O>::zeros(self.dim());
+        let mut stats = opencv::core::Mat::default();
+        let mut centroids = opencv::core::Mat::default();
+        let num_labels = opencv::imgproc::connected_components_with_stats(
+            self.as_image_mat()?.as_ref(),
+            labels.as_image_mat_mut()?.as_mut(),
+            &mut stats,
+            &mut centroids,
+            connectivity as i32,
+            O::as_cv_type(),
+        )
+        .change_context(NdCvError)?;
+        let stats = stats.as_ndarray()?.to_owned();
+        let centroids = centroids.as_ndarray()?.to_owned();
+        Ok(ConnectedComponentStats {
+            labels,
+            stats,
+            centroids,
+            num_labels,
+        })
+    }
+}
+
+// #[test]
+// fn test_connected_components() {
+//     use opencv::core::MatTrait as _;
+//     let mat = opencv::core::Mat::new_nd_with_default(&[10, 10], opencv::core::CV_8UC1, 0.into())
+//         .expect("failed");
+//     let roi1 = opencv::core::Rect::new(2, 2, 2, 2);
+//     let roi2 = opencv::core::Rect::new(6, 6, 3, 3);
+//     let mut mat1 = opencv::core::Mat::roi(&mat, roi1).expect("failed");
+//     mat1.set_scalar(1.into()).expect("failed");
+//     let mut mat2 = opencv::core::Mat::roi(&mat, roi2).expect("failed");
+//     mat2.set_scalar(1.into()).expect("failed");
+
+//     let array2: ndarray::ArrayView2<u8> = mat.as_ndarray().expect("failed");
+//     let output = array2
+//         .connected_components::<u16>(Connectivity::Four)
+//         .expect("failed");
+//     let expected = {
+//         let mut expected = ndarray::Array2::zeros((10, 10));
+//         expected.slice_mut(ndarray::s![2..4, 2..4]).fill(1);
+//         expected.slice_mut(ndarray::s![6..9, 6..9]).fill(2);
+//         expected
+//     };
+
+//     assert_eq!(output, expected);
+// }

ndcv-bridge/src/contours.rs

@@ -0,0 +1,270 @@
+//! <https://docs.rs/opencv/latest/opencv/imgproc/fn.find_contours.html>
+
+#![deny(warnings)]
+
+use crate::conversions::*;
+use crate::prelude_::*;
+use nalgebra::Point2;
+use ndarray::*;
+
+#[repr(C)]
+#[derive(Default, Debug, Copy, Clone, PartialEq, Eq)]
+pub enum ContourRetrievalMode {
+    #[default]
+    External = 0, // RETR_EXTERNAL
+    List = 1,      // RETR_LIST
+    CComp = 2,     // RETR_CCOMP
+    Tree = 3,      // RETR_TREE
+    FloodFill = 4, // RETR_FLOODFILL
+}
+
+#[repr(C)]
+#[derive(Default, Debug, Copy, Clone, PartialEq, Eq)]
+pub enum ContourApproximationMethod {
+    #[default]
+    None = 1, // CHAIN_APPROX_NONE
+    Simple = 2,   // CHAIN_APPROX_SIMPLE
+    Tc89L1 = 3,   // CHAIN_APPROX_TC89_L1
+    Tc89Kcos = 4, // CHAIN_APPROX_TC89_KCOS
+}
+
+#[derive(Debug, Clone)]
+pub struct ContourHierarchy {
+    pub next: i32,
+    pub previous: i32,
+    pub first_child: i32,
+    pub parent: i32,
+}
+
+#[derive(Debug, Clone)]
+pub struct ContourResult {
+    pub contours: Vec<Vec<Point2<i32>>>,
+    pub hierarchy: Vec<ContourHierarchy>,
+}
+
+mod seal {
+    pub trait Sealed {}
+    impl Sealed for u8 {}
+}
+
+pub trait NdCvFindContours<T: bytemuck::Pod + seal::Sealed>:
+    crate::image::NdImage + crate::conversions::NdAsImage<T, ndarray::Ix2>
+{
+    fn find_contours(
+        &self,
+        mode: ContourRetrievalMode,
+        method: ContourApproximationMethod,
+    ) -> Result<Vec<Vec<Point2<i32>>>, NdCvError>;
+
+    fn find_contours_with_hierarchy(
+        &self,
+        mode: ContourRetrievalMode,
+        method: ContourApproximationMethod,
+    ) -> Result<ContourResult, NdCvError>;
+
+    fn find_contours_def(&self) -> Result<Vec<Vec<Point2<i32>>>, NdCvError> {
+        self.find_contours(
+            ContourRetrievalMode::External,
+            ContourApproximationMethod::Simple,
+        )
+    }
+
+    fn find_contours_with_hierarchy_def(&self) -> Result<ContourResult, NdCvError> {
+        self.find_contours_with_hierarchy(
+            ContourRetrievalMode::External,
+            ContourApproximationMethod::Simple,
+        )
+    }
+}
+
+pub trait NdCvContourArea<T: bytemuck::Pod> {
+    fn contours_area(&self, oriented: bool) -> Result<f64, NdCvError>;
+
+    fn contours_area_def(&self) -> Result<f64, NdCvError> {
+        self.contours_area(false)
+    }
+}
+
+impl<T: ndarray::RawData + ndarray::Data<Elem = u8>> NdCvFindContours<u8> for ArrayBase<T, Ix2> {
+    fn find_contours(
+        &self,
+        mode: ContourRetrievalMode,
+        method: ContourApproximationMethod,
+    ) -> Result<Vec<Vec<Point2<i32>>>, NdCvError> {
+        let cv_self = self.as_image_mat()?;
+        let mut contours = opencv::core::Vector::<opencv::core::Vector<opencv::core::Point>>::new();
+
+        opencv::imgproc::find_contours(
+            &*cv_self,
+            &mut contours,
+            mode as i32,
+            method as i32,
+            opencv::core::Point::new(0, 0),
+        )
+        .change_context(NdCvError)
+        .attach_printable("Failed to find contours")?;
+        let mut result: Vec<Vec<Point2<i32>>> = Vec::new();
+
+        for i in 0..contours.len() {
+            let contour = contours.get(i).change_context(NdCvError)?;
+            let points: Vec<Point2<i32>> =
+                contour.iter().map(|pt| Point2::new(pt.x, pt.y)).collect();
+            result.push(points);
+        }
+
+        Ok(result)
+    }
+
+    fn find_contours_with_hierarchy(
+        &self,
+        mode: ContourRetrievalMode,
+        method: ContourApproximationMethod,
+    ) -> Result<ContourResult, NdCvError> {
+        let cv_self = self.as_image_mat()?;
+        let mut contours = opencv::core::Vector::<opencv::core::Vector<opencv::core::Point>>::new();
+        let mut hierarchy = opencv::core::Vector::<opencv::core::Vec4i>::new();
+
+        opencv::imgproc::find_contours_with_hierarchy(
+            &*cv_self,
+            &mut contours,
+            &mut hierarchy,
+            mode as i32,
+            method as i32,
+            opencv::core::Point::new(0, 0),
+        )
+        .change_context(NdCvError)
+        .attach_printable("Failed to find contours with hierarchy")?;
+        let mut contour_list: Vec<Vec<Point2<i32>>> = Vec::new();
+
+        for i in 0..contours.len() {
+            let contour = contours.get(i).change_context(NdCvError)?;
+            let points: Vec<Point2<i32>> =
+                contour.iter().map(|pt| Point2::new(pt.x, pt.y)).collect();
+            contour_list.push(points);
+        }
+
+        let mut hierarchy_list = Vec::new();
+        for i in 0..hierarchy.len() {
+            let h = hierarchy.get(i).change_context(NdCvError)?;
+            hierarchy_list.push(ContourHierarchy {
+                next: h[0],
+                previous: h[1],
+                first_child: h[2],
+                parent: h[3],
+            });
+        }
+
+        Ok(ContourResult {
+            contours: contour_list,
+            hierarchy: hierarchy_list,
+        })
+    }
+}
+
+impl<T> NdCvContourArea<T> for Vec<Point2<T>>
+where
+    T: bytemuck::Pod + num::traits::AsPrimitive<i32> + std::cmp::PartialEq + std::fmt::Debug + Copy,
+{
+    fn contours_area(&self, oriented: bool) -> Result<f64, NdCvError> {
+        if self.is_empty() {
+            return Ok(0.0);
+        }
+
+        let mut cv_contour: opencv::core::Vector<opencv::core::Point> = opencv::core::Vector::new();
+        self.iter().for_each(|point| {
+            cv_contour.push(opencv::core::Point::new(
+                point.coords[0].as_(),
+                point.coords[1].as_(),
+            ));
+        });
+
+        opencv::imgproc::contour_area(&cv_contour, oriented)
+            .change_context(NdCvError)
+            .attach_printable("Failed to calculate contour area")
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use ndarray::Array2;
+
+    fn simple_binary_rect_image() -> Array2<u8> {
+        let mut img = Array2::<u8>::zeros((10, 10));
+        for i in 2..8 {
+            for j in 3..7 {
+                img[(i, j)] = 255;
+            }
+        }
+        img
+    }
+
+    #[test]
+    fn test_find_contours_external_simple() {
+        let img = simple_binary_rect_image();
+        let contours = img
+            .find_contours(
+                ContourRetrievalMode::External,
+                ContourApproximationMethod::Simple,
+            )
+            .expect("Failed to find contours");
+        assert_eq!(contours.len(), 1);
+        assert!(contours[0].len() >= 4);
+    }
+
+    #[test]
+    fn test_find_contours_with_hierarchy() {
+        let img = simple_binary_rect_image();
+        let res = img
+            .find_contours_with_hierarchy(
+                ContourRetrievalMode::External,
+                ContourApproximationMethod::Simple,
+            )
+            .expect("Failed to find contours with hierarchy");
+        assert_eq!(res.contours.len(), 1);
+        assert_eq!(res.hierarchy.len(), 1);
+
+        let h = &res.hierarchy[0];
+        assert_eq!(h.parent, -1);
+        assert_eq!(h.first_child, -1);
+    }
+
+    #[test]
+    fn test_default_methods() {
+        let img = simple_binary_rect_image();
+        let contours = img.find_contours_def().unwrap();
+        let res = img.find_contours_with_hierarchy_def().unwrap();
+        assert_eq!(contours.len(), 1);
+        assert_eq!(res.contours.len(), 1);
+    }
+
+    #[test]
+    fn test_contour_area_calculation() {
+        let img = simple_binary_rect_image();
+        let contours = img.find_contours_def().unwrap();
+        let expected_area = 15.;
+        let area = contours[0].contours_area_def().unwrap();
+        assert!(
+            (area - expected_area).abs() < 1.0,
+            "Area mismatch: got {area}, expected {expected_area}",
+        );
+    }
+
+    #[test]
+    fn test_empty_input_returns_no_contours() {
+        let img = Array2::<u8>::zeros((10, 10));
+        let contours = img.find_contours_def().unwrap();
+        assert!(contours.is_empty());
+
+        let res = img.find_contours_with_hierarchy_def().unwrap();
+        assert!(res.contours.is_empty());
+        assert!(res.hierarchy.is_empty());
+    }
+
+    #[test]
+    fn test_contour_area_empty_contour() {
+        let contour: Vec<Point2<i32>> = vec![];
+        let area = contour.contours_area_def().unwrap();
+        assert_eq!(area, 0.0);
+    }
+}

ndcv-bridge/src/conversions.rs

@@ -0,0 +1,337 @@
+//! Mat <--> ndarray conversion traits
+//!
+//! Conversion Table
+//!
+//! | ndarray           | Mat     |
+//! |---------          |-----    |
+//! | Array<T, Ix1>     | Mat(ndims = 1, channels = 1)   |
+//! | Array<T, Ix2>     | Mat(ndims = 2, channels = 1)   |
+//! | Array<T, Ix2>     | Mat(ndims = 1, channels = X)   |
+//! | Array<T, Ix3>     | Mat(ndims = 3, channels = 1)   |
+//! | Array<T, Ix3>     | Mat(ndims = 2, channels = X)   |
+//! | Array<T, Ix4>     | Mat(ndims = 4, channels = 1)   |
+//! | Array<T, Ix4>     | Mat(ndims = 3, channels = X)   |
+//! | Array<T, Ix5>     | Mat(ndims = 5, channels = 1)   |
+//! | Array<T, Ix5>     | Mat(ndims = 4, channels = X)   |
+//! | Array<T, Ix6>     | Mat(ndims = 6, channels = 1)   |
+//! | Array<T, Ix6>     | Mat(ndims = 5, channels = X)   |
+//!
+//! // X is the last dimension
+use crate::NdCvError;
+use crate::type_depth;
+use error_stack::*;
+use ndarray::{Ix2, Ix3};
+use opencv::core::MatTraitConst;
+mod impls;
+pub(crate) mod matref;
+use matref::{MatRef, MatRefMut};
+
+pub(crate) mod seal {
+    pub trait SealedInternal {}
+    impl<T, S: ndarray::Data<Elem = T>, D> SealedInternal for ndarray::ArrayBase<S, D> {}
+    // impl<T, S: ndarray::DataMut<Elem = T>, D> SealedInternal for ndarray::ArrayBase<S, D> {}
+}
+
+pub trait NdCvConversion<T: bytemuck::Pod + Copy, D: ndarray::Dimension>:
+    seal::SealedInternal + Sized
+{
+    fn to_mat(&self) -> Result<opencv::core::Mat, NdCvError>;
+    fn from_mat(
+        mat: opencv::core::Mat,
+    ) -> Result<ndarray::ArrayBase<ndarray::OwnedRepr<T>, D>, NdCvError>;
+}
+
+impl<T: bytemuck::Pod + Copy, S: ndarray::Data<Elem = T>, D: ndarray::Dimension>
+    NdCvConversion<T, D> for ndarray::ArrayBase<S, D>
+where
+    Self: NdAsImage<T, D>,
+{
+    fn to_mat(&self) -> Result<opencv::core::Mat, NdCvError> {
+        Ok(self.as_image_mat()?.mat.clone())
+    }
+
+    fn from_mat(
+        mat: opencv::core::Mat,
+    ) -> Result<ndarray::ArrayBase<ndarray::OwnedRepr<T>, D>, NdCvError> {
+        let ndarray = unsafe { impls::mat_to_ndarray::<T, D>(&mat) }.change_context(NdCvError)?;
+        Ok(ndarray.to_owned())
+    }
+}
+
+pub trait MatAsNd {
+    fn as_ndarray<T: bytemuck::Pod, D: ndarray::Dimension>(
+        &self,
+    ) -> Result<ndarray::ArrayView<T, D>, NdCvError>;
+}
+
+impl MatAsNd for opencv::core::Mat {
+    fn as_ndarray<T: bytemuck::Pod, D: ndarray::Dimension>(
+        &self,
+    ) -> Result<ndarray::ArrayView<T, D>, NdCvError> {
+        unsafe { impls::mat_to_ndarray::<T, D>(self) }.change_context(NdCvError)
+    }
+}
+
+pub trait NdAsMat<T: bytemuck::Pod + Copy, D: ndarray::Dimension> {
+    fn as_single_channel_mat(&self) -> Result<MatRef, NdCvError>;
+    fn as_multi_channel_mat(&self) -> Result<MatRef, NdCvError>;
+}
+
+pub trait NdAsMatMut<T: bytemuck::Pod + Copy, D: ndarray::Dimension>: NdAsMat<T, D> {
+    fn as_single_channel_mat_mut(&mut self) -> Result<MatRefMut, NdCvError>;
+    fn as_multi_channel_mat_mut(&mut self) -> Result<MatRefMut, NdCvError>;
+}
+
+impl<T: bytemuck::Pod, S: ndarray::Data<Elem = T>, D: ndarray::Dimension> NdAsMat<T, D>
+    for ndarray::ArrayBase<S, D>
+{
+    fn as_single_channel_mat(&self) -> Result<MatRef, NdCvError> {
+        let mat = unsafe { impls::ndarray_to_mat_regular(self) }.change_context(NdCvError)?;
+        Ok(MatRef::new(mat))
+    }
+    fn as_multi_channel_mat(&self) -> Result<MatRef, NdCvError> {
+        let mat = unsafe { impls::ndarray_to_mat_consolidated(self) }.change_context(NdCvError)?;
+        Ok(MatRef::new(mat))
+    }
+}
+
+impl<T: bytemuck::Pod, S: ndarray::DataMut<Elem = T>, D: ndarray::Dimension> NdAsMatMut<T, D>
+    for ndarray::ArrayBase<S, D>
+{
+    fn as_single_channel_mat_mut(&mut self) -> Result<MatRefMut, NdCvError> {
+        let mat = unsafe { impls::ndarray_to_mat_regular(self) }.change_context(NdCvError)?;
+        Ok(MatRefMut::new(mat))
+    }
+
+    fn as_multi_channel_mat_mut(&mut self) -> Result<MatRefMut, NdCvError> {
+        let mat = unsafe { impls::ndarray_to_mat_consolidated(self) }.change_context(NdCvError)?;
+        Ok(MatRefMut::new(mat))
+    }
+}
+
+pub trait NdAsImage<T: bytemuck::Pod, D: ndarray::Dimension> {
+    fn as_image_mat(&self) -> Result<MatRef, NdCvError>;
+}
+
+pub trait NdAsImageMut<T: bytemuck::Pod, D: ndarray::Dimension> {
+    fn as_image_mat_mut(&mut self) -> Result<MatRefMut, NdCvError>;
+}
+
+impl<T, S> NdAsImage<T, Ix2> for ndarray::ArrayBase<S, Ix2>
+where
+    T: bytemuck::Pod + Copy,
+    S: ndarray::Data<Elem = T>,
+{
+    fn as_image_mat(&self) -> Result<MatRef, NdCvError> {
+        self.as_single_channel_mat()
+    }
+}
+
+impl<T, S> NdAsImageMut<T, Ix2> for ndarray::ArrayBase<S, Ix2>
+where
+    T: bytemuck::Pod + Copy,
+    S: ndarray::DataMut<Elem = T>,
+{
+    fn as_image_mat_mut(&mut self) -> Result<MatRefMut, NdCvError> {
+        self.as_single_channel_mat_mut()
+    }
+}
+
+impl<T, S> NdAsImage<T, Ix3> for ndarray::ArrayBase<S, Ix3>
+where
+    T: bytemuck::Pod + Copy,
+    S: ndarray::Data<Elem = T>,
+{
+    fn as_image_mat(&self) -> Result<MatRef, NdCvError> {
+        self.as_multi_channel_mat()
+    }
+}
+
+impl<T, S> NdAsImageMut<T, Ix3> for ndarray::ArrayBase<S, Ix3>
+where
+    T: bytemuck::Pod + Copy,
+    S: ndarray::DataMut<Elem = T>,
+{
+    fn as_image_mat_mut(&mut self) -> Result<MatRefMut, NdCvError> {
+        self.as_multi_channel_mat_mut()
+    }
+}
+
+// #[test]
+// fn test_1d_mat_to_ndarray() {
+//     let mat = opencv::core::Mat::new_nd_with_default(
+//         &[10],
+//         opencv::core::CV_MAKE_TYPE(opencv::core::CV_8U, 1),
+//         200.into(),
+//     )
+//     .expect("failed");
+//     let array: ndarray::ArrayView1<u8> = mat.as_ndarray().expect("failed");
+//     array.into_iter().for_each(|&x| assert_eq!(x, 200));
+// }
+
+// #[test]
+// fn test_2d_mat_to_ndarray() {
+//     let mat = opencv::core::Mat::new_nd_with_default(
+//         &[10],
+//         opencv::core::CV_16SC3,
+//         (200, 200, 200).into(),
+//     )
+//     .expect("failed");
+//     let array2: ndarray::ArrayView2<i16> = mat.as_ndarray().expect("failed");
+//     assert_eq!(array2.shape(), [10, 3]);
+//     array2.into_iter().for_each(|&x| {
+//         assert_eq!(x, 200);
+//     });
+// }
+
+// #[test]
+// fn test_3d_mat_to_ndarray() {
+//     let mat = opencv::core::Mat::new_nd_with_default(
+//         &[20, 30],
+//         opencv::core::CV_32FC3,
+//         (200, 200, 200).into(),
+//     )
+//     .expect("failed");
+//     let array2: ndarray::ArrayView3<f32> = mat.as_ndarray().expect("failed");
+//     array2.into_iter().for_each(|&x| {
+//         assert_eq!(x, 200f32);
+//     });
+// }
+
+// #[test]
+// fn test_mat_to_dyn_ndarray() {
+//     let mat = opencv::core::Mat::new_nd_with_default(&[10], opencv::core::CV_8UC1, 200.into())
+//         .expect("failed");
+//     let array2: ndarray::ArrayViewD<u8> = mat.as_ndarray().expect("failed");
+//     array2.into_iter().for_each(|&x| assert_eq!(x, 200));
+// }
+
+// #[test]
+// fn test_3d_mat_to_ndarray_4k() {
+//     let mat = opencv::core::Mat::new_nd_with_default(
+//         &[4096, 4096],
+//         opencv::core::CV_8UC3,
+//         (255, 0, 255).into(),
+//     )
+//     .expect("failed");
+//     let array2: ndarray::ArrayView3<u8> = (mat).as_ndarray().expect("failed");
+//     array2.exact_chunks((1, 1, 3)).into_iter().for_each(|x| {
+//         assert_eq!(x[(0, 0, 0)], 255);
+//         assert_eq!(x[(0, 0, 1)], 0);
+//         assert_eq!(x[(0, 0, 2)], 255);
+//     });
+// }
+
+// // #[test]
+// // fn test_3d_mat_to_ndarray_8k() {
+// //     let mat = opencv::core::Mat::new_nd_with_default(
+// //         &[8192, 8192],
+// //         opencv::core::CV_8UC3,
+// //         (255, 0, 255).into(),
+// //     )
+// //     .expect("failed");
+// //     let array2 = ndarray::Array3::<u8>::from_mat(mat).expect("failed");
+// //     array2.exact_chunks((1, 1, 3)).into_iter().for_each(|x| {
+// //         assert_eq!(x[(0, 0, 0)], 255);
+// //         assert_eq!(x[(0, 0, 1)], 0);
+// //         assert_eq!(x[(0, 0, 2)], 255);
+// //     });
+// // }
+
+// #[test]
+// pub fn test_mat_to_nd_default_strides() {
+//     let mat = opencv::core::Mat::new_rows_cols_with_default(
+//         10,
+//         10,
+//         opencv::core::CV_8UC3,
+//         opencv::core::VecN([10f64, 0.0, 0.0, 0.0]),
+//     )
+//     .expect("failed");
+//     let array = unsafe { impls::mat_to_ndarray::<u8, Ix3>(&mat) }.expect("failed");
+//     assert_eq!(array.shape(), [10, 10, 3]);
+//     assert_eq!(array.strides(), [30, 3, 1]);
+//     assert_eq!(array[(0, 0, 0)], 10);
+// }
+
+// #[test]
+// pub fn test_mat_to_nd_custom_strides() {
+//     let mat = opencv::core::Mat::new_rows_cols_with_default(
+//         10,
+//         10,
+//         opencv::core::CV_8UC3,
+//         opencv::core::VecN([10f64, 0.0, 0.0, 0.0]),
+//     )
+//     .unwrap();
+//     let mat_roi = opencv::core::Mat::roi(&mat, opencv::core::Rect::new(3, 2, 3, 5))
+//         .expect("failed to get roi");
+//     let array = unsafe { impls::mat_to_ndarray::<u8, Ix3>(&mat_roi) }.expect("failed");
+//     assert_eq!(array.shape(), [5, 3, 3]);
+//     assert_eq!(array.strides(), [30, 3, 1]);
+//     assert_eq!(array[(0, 0, 0)], 10);
+// }
+
+// #[test]
+// pub fn test_non_continuous_3d() {
+//     let array = ndarray::Array3::<f32>::from_shape_fn((10, 10, 4), |(i, j, k)| {
+//         ((i + 1) * (j + 1) * (k + 1)) as f32
+//     });
+//     let slice = array.slice(ndarray::s![3..7, 3..7, 0..4]);
+//     let mat = unsafe { impls::ndarray_to_mat_consolidated(&slice) }.unwrap();
+//     let arr = unsafe { impls::mat_to_ndarray::<f32, Ix3>(&mat).unwrap() };
+//     assert!(slice == arr);
+// }
+
+// #[test]
+// pub fn test_5d_array() {
+//     let array = ndarray::Array5::<f32>::ones((1, 2, 3, 4, 5));
+//     let mat = unsafe { impls::ndarray_to_mat_consolidated(&array) }.unwrap();
+//     let arr = unsafe { impls::mat_to_ndarray::<f32, ndarray::Ix5>(&mat).unwrap() };
+//     assert_eq!(array, arr);
+// }
+
+// #[test]
+// pub fn test_3d_array() {
+//     let array = ndarray::Array3::<f32>::ones((23, 31, 33));
+//     let mat = unsafe { impls::ndarray_to_mat_consolidated(&array) }.unwrap();
+//     let arr = unsafe { impls::mat_to_ndarray::<f32, ndarray::Ix3>(&mat).unwrap() };
+//     assert_eq!(array, arr);
+// }
+
+// #[test]
+// pub fn test_2d_array() {
+//     let array = ndarray::Array2::<f32>::ones((23, 31));
+//     let mat = unsafe { impls::ndarray_to_mat_consolidated(&array) }.unwrap();
+//     let arr = unsafe { impls::mat_to_ndarray::<f32, ndarray::Ix2>(&mat).unwrap() };
+//     assert_eq!(array, arr);
+// }
+
+// #[test]
+// #[should_panic]
+// pub fn test_1d_array_consolidated() {
+//     let array = ndarray::Array1::<f32>::ones(23);
+//     let mat = unsafe { impls::ndarray_to_mat_consolidated(&array) }.unwrap();
+//     let arr = unsafe { impls::mat_to_ndarray::<f32, ndarray::Ix1>(&mat).unwrap() };
+//     assert_eq!(array, arr);
+// }
+
+// #[test]
+// pub fn test_1d_array_regular() {
+//     let array = ndarray::Array1::<f32>::ones(23);
+//     let mat = unsafe { impls::ndarray_to_mat_regular(&array) }.unwrap();
+//     let arr = unsafe { impls::mat_to_ndarray::<f32, ndarray::Ix1>(&mat).unwrap() };
+//     assert_eq!(array, arr);
+// }
+
+// #[test]
+// pub fn test_2d_array_regular() {
+//     let array = ndarray::Array2::<f32>::ones((23, 31));
+//     let mat = unsafe { impls::ndarray_to_mat_regular(&array) }.unwrap();
+//     let arr = unsafe { impls::mat_to_ndarray::<f32, ndarray::Ix2>(&mat).unwrap() };
+//     assert_eq!(array, arr);
+// }
+
+// #[test]
+// pub fn test_ndcv_1024_1024_to_mat() {
+//     let array = ndarray::Array2::<f32>::ones((1024, 1024));
+//     let _mat = array.to_mat().unwrap();
+// }

ndcv-bridge/src/conversions/impls.rs

@@ -0,0 +1,168 @@
+use super::*;
+use core::ffi::*;
+use opencv::core::prelude::*;
+pub(crate) unsafe fn ndarray_to_mat_regular<
+    T,
+    S: ndarray::Data<Elem = T>,
+    D: ndarray::Dimension,
+>(
+    input: &ndarray::ArrayBase<S, D>,
+) -> Result<opencv::core::Mat, NdCvError> {
+    let shape = input.shape();
+    let strides = input.strides();
+
+    // let channels = shape.last().copied().unwrap_or(1);
+    // if channels > opencv::core::CV_CN_MAX as usize {
+    //     Err(Report::new(NdCvError).attach_printable(format!(
+    //             "Number of channels({channels}) exceeds CV_CN_MAX({}) use the regular version of the function", opencv::core::CV_CN_MAX
+    //         )))?;
+    // }
+
+    // let size_len = shape.len();
+    let size = shape.iter().copied().map(|f| f as i32).collect::<Vec<_>>();
+    // Step len for ndarray is always 1 less than ndims
+    let step_len = strides.len() - 1;
+    let step = strides
+        .iter()
+        .take(step_len)
+        .copied()
+        .map(|f| f as usize * core::mem::size_of::<T>())
+        .collect::<Vec<_>>();
+
+    let data_ptr = input.as_ptr() as *const c_void;
+
+    let typ = opencv::core::CV_MAKETYPE(type_depth::<T>(), 1);
+    let mat = opencv::core::Mat::new_nd_with_data_unsafe(
+        size.as_slice(),
+        typ,
+        data_ptr.cast_mut(),
+        Some(step.as_slice()),
+    )
+    .change_context(NdCvError)?;
+
+    Ok(mat)
+}
+
+pub(crate) unsafe fn ndarray_to_mat_consolidated<
+    T,
+    S: ndarray::Data<Elem = T>,
+    D: ndarray::Dimension,
+>(
+    input: &ndarray::ArrayBase<S, D>,
+) -> Result<opencv::core::Mat, NdCvError> {
+    let shape = input.shape();
+    let strides = input.strides();
+
+    let channels = shape.last().copied().unwrap_or(1);
+    if channels > opencv::core::CV_CN_MAX as usize {
+        Err(Report::new(NdCvError).attach_printable(format!(
+                "Number of channels({channels}) exceeds CV_CN_MAX({}) use the regular version of the function", opencv::core::CV_CN_MAX
+            )))?;
+    }
+
+    if shape.len() > 2 {
+        // Basically the second last stride is used to jump from one column to next
+        // But opencv only keeps ndims - 1 strides so we can't have the column stride as that
+        // will be lost
+        if shape.last() != strides.get(strides.len() - 2).map(|x| *x as usize).as_ref() {
+            Err(Report::new(NdCvError).attach_printable(
+                "You cannot slice into the last axis in ndarray when converting to mat",
+            ))?;
+        }
+    } else if shape.len() == 1 {
+        return Err(Report::new(NdCvError).attach_printable(
+            "You cannot convert a 1D array to a Mat while using the consolidated version",
+        ));
+    }
+
+    // Since this is the consolidated version we should always only have ndims - 1 sizes and
+    // ndims - 2 strides
+
+    let size_len = shape.len() - 1; // Since we move last axis into the channel
+    let size = shape
+        .iter()
+        .take(size_len)
+        .map(|f| *f as i32)
+        .collect::<Vec<_>>();
+
+    let step_len = strides.len() - 1;
+    let step = strides
+        .iter()
+        .take(step_len)
+        .map(|f| *f as usize * core::mem::size_of::<T>())
+        .collect::<Vec<_>>();
+
+    let data_ptr = input.as_ptr() as *const c_void;
+
+    let typ = opencv::core::CV_MAKETYPE(type_depth::<T>(), channels as i32);
+
+    let mat = opencv::core::Mat::new_nd_with_data_unsafe(
+        size.as_slice(),
+        typ,
+        data_ptr.cast_mut(),
+        Some(step.as_slice()),
+    )
+    .change_context(NdCvError)?;
+
+    Ok(mat)
+}
+
+pub(crate) unsafe fn mat_to_ndarray<T: bytemuck::Pod, D: ndarray::Dimension>(
+    mat: &opencv::core::Mat,
+) -> Result<ndarray::ArrayView<'_, T, D>, NdCvError> {
+    let depth = mat.depth();
+    if type_depth::<T>() != depth {
+        return Err(Report::new(NdCvError).attach_printable(format!(
+            "Expected type Mat<{}> ({}), got Mat<{}> ({})",
+            std::any::type_name::<T>(),
+            type_depth::<T>(),
+            crate::depth_type(depth),
+            depth,
+        )));
+    }
+
+    // Since a dims always returns >= 2 we can't use this to check if it's a 1D array
+    // So we compare the first axis to the total to see if its a 1D array
+    let is_1d = mat.total() as i32 == mat.rows();
+    let dims = is_1d.then_some(1).unwrap_or(mat.dims());
+    let channels = mat.channels();
+    let ndarray_size = (channels != 1).then_some(dims + 1).unwrap_or(dims) as usize;
+    if let Some(ndim) = D::NDIM {
+        // When channels is not 1,
+        // the last dimension is the channels
+        // Array1 -> Mat(ndims = 1, channels = 1)
+        // Array2 -> Mat(ndims = 1, channels = X)
+        // Array2 -> Mat(ndims = 2, channels = 1)
+        // Array3 -> Mat(ndims = 2, channels = X)
+        // Array3 -> Mat(ndims = 3, channels = 1)
+        // ...
+        if ndim != dims as usize && channels == 1 {
+            return Err(Report::new(NdCvError)
+                .attach_printable(format!("Expected {}D array, got {}D", ndim, ndarray_size)));
+        }
+    }
+
+    let mat_size = mat.mat_size();
+    let sizes = (0..dims)
+        .map(|i| mat_size.get(i).change_context(NdCvError))
+        .chain([Ok(channels)])
+        .map(|x| x.map(|x| x as usize))
+        .take(ndarray_size)
+        .collect::<Result<Vec<_>, NdCvError>>()
+        .change_context(NdCvError)?;
+    let strides = (0..(dims - 1))
+        .map(|i| mat.step1(i).change_context(NdCvError))
+        .chain([
+            Ok(channels as usize),
+            Ok((channels == 1).then_some(0).unwrap_or(1)),
+        ])
+        .take(ndarray_size)
+        .collect::<Result<Vec<_>, NdCvError>>()
+        .change_context(NdCvError)?;
+    use ndarray::ShapeBuilder;
+    let shape = sizes.strides(strides);
+    let raw_array = ndarray::RawArrayView::from_shape_ptr(shape, mat.data() as *const T)
+        .into_dimensionality()
+        .change_context(NdCvError)?;
+    Ok(unsafe { raw_array.deref_into_view() })
+}

ndcv-bridge/src/conversions/matref.rs

@@ -0,0 +1,73 @@
+#[derive(Debug, Clone)]
+pub struct MatRef<'a> {
+    pub(crate) mat: opencv::core::Mat,
+    pub(crate) _marker: core::marker::PhantomData<&'a ()>,
+}
+
+impl MatRef<'_> {
+    pub fn clone_pointee(&self) -> opencv::core::Mat {
+        self.mat.clone()
+    }
+}
+
+impl MatRef<'_> {
+    pub fn new<'a>(mat: opencv::core::Mat) -> MatRef<'a> {
+        MatRef {
+            mat,
+            _marker: core::marker::PhantomData,
+        }
+    }
+}
+
+impl AsRef<opencv::core::Mat> for MatRef<'_> {
+    fn as_ref(&self) -> &opencv::core::Mat {
+        &self.mat
+    }
+}
+
+impl AsRef<opencv::core::Mat> for MatRefMut<'_> {
+    fn as_ref(&self) -> &opencv::core::Mat {
+        &self.mat
+    }
+}
+
+impl AsMut<opencv::core::Mat> for MatRefMut<'_> {
+    fn as_mut(&mut self) -> &mut opencv::core::Mat {
+        &mut self.mat
+    }
+}
+
+#[derive(Debug, Clone)]
+pub struct MatRefMut<'a> {
+    pub(crate) mat: opencv::core::Mat,
+    pub(crate) _marker: core::marker::PhantomData<&'a mut ()>,
+}
+
+impl MatRefMut<'_> {
+    pub fn new<'a>(mat: opencv::core::Mat) -> MatRefMut<'a> {
+        MatRefMut {
+            mat,
+            _marker: core::marker::PhantomData,
+        }
+    }
+}
+
+impl core::ops::Deref for MatRef<'_> {
+    type Target = opencv::core::Mat;
+    fn deref(&self) -> &Self::Target {
+        &self.mat
+    }
+}
+
+impl core::ops::Deref for MatRefMut<'_> {
+    type Target = opencv::core::Mat;
+    fn deref(&self) -> &Self::Target {
+        &self.mat
+    }
+}
+
+impl core::ops::DerefMut for MatRefMut<'_> {
+    fn deref_mut(&mut self) -> &mut Self::Target {
+        &mut self.mat
+    }
+}

ndcv-bridge/src/fir.rs

@@ -0,0 +1,262 @@
+use error_stack::*;
+use fast_image_resize::*;
+use images::{Image, ImageRef};
+#[derive(Debug, Clone, thiserror::Error)]
+#[error("NdFirError")]
+pub struct NdFirError;
+type Result<T, E = Report<NdFirError>> = std::result::Result<T, E>;
+
+pub trait NdAsImage<T: seal::Sealed, D: ndarray::Dimension>: Sized {
+    fn as_image_ref(&self) -> Result<ImageRef<'_>>;
+}
+
+pub trait NdAsImageMut<T: seal::Sealed, D: ndarray::Dimension>: Sized {
+    fn as_image_ref_mut(&mut self) -> Result<Image<'_>>;
+}
+
+pub struct NdarrayImageContainer<'a, T: seal::Sealed, D: ndarray::Dimension> {
+    #[allow(dead_code)]
+    data: ndarray::ArrayView<'a, T, D>,
+    pub _phantom: std::marker::PhantomData<(T, D)>,
+}
+
+impl<'a, T: seal::Sealed> NdarrayImageContainer<'a, T, ndarray::Ix3> {
+    pub fn new<S: ndarray::Data<Elem = T>>(array: &'a ndarray::ArrayBase<S, ndarray::Ix3>) -> Self {
+        Self {
+            data: array.view(),
+            _phantom: std::marker::PhantomData,
+        }
+    }
+}
+
+impl<'a, T: seal::Sealed> NdarrayImageContainer<'a, T, ndarray::Ix2> {
+    pub fn new<S: ndarray::Data<Elem = T>>(array: &'a ndarray::ArrayBase<S, ndarray::Ix2>) -> Self {
+        Self {
+            data: array.view(),
+            _phantom: std::marker::PhantomData,
+        }
+    }
+}
+pub struct NdarrayImageContainerMut<'a, T: seal::Sealed, D: ndarray::Dimension> {
+    #[allow(dead_code)]
+    data: ndarray::ArrayViewMut<'a, T, D>,
+}
+
+impl<'a, T: seal::Sealed> NdarrayImageContainerMut<'a, T, ndarray::Ix3> {
+    pub fn new<S: ndarray::DataMut<Elem = T>>(
+        array: &'a mut ndarray::ArrayBase<S, ndarray::Ix3>,
+    ) -> Self {
+        Self {
+            data: array.view_mut(),
+        }
+    }
+}
+
+impl<'a, T: seal::Sealed> NdarrayImageContainerMut<'a, T, ndarray::Ix2> {
+    pub fn new<S: ndarray::DataMut<Elem = T>>(
+        array: &'a mut ndarray::ArrayBase<S, ndarray::Ix2>,
+    ) -> Self {
+        Self {
+            data: array.view_mut(),
+        }
+    }
+}
+
+pub struct NdarrayImageContainerTyped<'a, T: seal::Sealed, D: ndarray::Dimension, P: PixelTrait> {
+    #[allow(dead_code)]
+    data: ndarray::ArrayView<'a, T, D>,
+    __marker: std::marker::PhantomData<P>,
+}
+
+// unsafe impl<'a, T: seal::Sealed + Sync + InnerPixel, P: PixelTrait> ImageView
+//     for NdarrayImageContainerTyped<'a, T, ndarray::Ix3, P>
+// where
+//     T: bytemuck::Pod,
+// {
+//     type Pixel = P;
+//     fn width(&self) -> u32 {
+//         self.data.shape()[1] as u32
+//     }
+//     fn height(&self) -> u32 {
+//         self.data.shape()[0] as u32
+//     }
+//     fn iter_rows(&self, start_row: u32) -> impl Iterator<Item = &[Self::Pixel]> {
+//         self.data
+//             .rows()
+//             .into_iter()
+//             .skip(start_row as usize)
+//             .map(|row| {
+//                 row.as_slice()
+//                     .unwrap_or_default()
+//                     .chunks_exact(P::CHANNELS as usize)
+//             })
+//     }
+// }
+
+// impl<'a, T: fast_image_resize::pixels::InnerPixel + seal::Sealed, D: ndarray::Dimension>
+//     fast_image_resize::IntoImageView for NdarrayImageContainer<'a, T, D>
+// {
+//     fn pixel_type(&self) -> Option<PixelType> {
+//         match D::NDIM {
+//             Some(2) => Some(to_pixel_type::<T>(1).expect("Failed to convert to pixel type")),
+//             Some(3) => Some(
+//                 to_pixel_type::<T>(self.data.shape()[2]).expect("Failed to convert to pixel type"),
+//             ),
+//             _ => None,
+//         }
+//     }
+//     fn width(&self) -> u32 {
+//         self.data.shape()[1] as u32
+//     }
+//     fn height(&self) -> u32 {
+//         self.data.shape()[0] as u32
+//     }
+//     fn image_view<P: PixelTrait>(&'a self) -> Option<NdarrayImageContainerTyped<'a, T, D, P>> {
+//         Some(NdarrayImageContainerTyped {
+//             data: self.data.view(),
+//             __marker: std::marker::PhantomData,
+//         })
+//     }
+// }
+
+pub fn to_pixel_type<T: seal::Sealed>(u: usize) -> Result<PixelType> {
+    match (core::any::type_name::<T>(), u) {
+        ("u8", 1) => Ok(PixelType::U8),
+        ("u8", 2) => Ok(PixelType::U8x2),
+        ("u8", 3) => Ok(PixelType::U8x3),
+        ("u8", 4) => Ok(PixelType::U8x4),
+        ("u16", 1) => Ok(PixelType::U16),
+        ("i32", 1) => Ok(PixelType::I32),
+        ("f32", 1) => Ok(PixelType::F32),
+        ("f32", 2) => Ok(PixelType::F32x2),
+        ("f32", 3) => Ok(PixelType::F32x3),
+        ("f32", 4) => Ok(PixelType::F32x4),
+        _ => Err(Report::new(NdFirError).attach_printable("Unsupported pixel type")),
+    }
+}
+
+mod seal {
+    pub trait Sealed {}
+    impl Sealed for u8 {}
+    impl Sealed for u16 {}
+    impl Sealed for i32 {}
+    impl Sealed for f32 {}
+}
+
+impl<S: ndarray::Data<Elem = T>, T: seal::Sealed + bytemuck::Pod, D: ndarray::Dimension>
+    NdAsImage<T, D> for ndarray::ArrayBase<S, D>
+{
+    /// Clones self and makes a new image
+    fn as_image_ref(&self) -> Result<ImageRef> {
+        let shape = self.shape();
+        let rows = *shape
+            .first()
+            .ok_or_else(|| Report::new(NdFirError).attach_printable("Failed to get rows"))?
+            as u32;
+        let cols = *shape.get(1).unwrap_or(&1) as u32;
+        let channels = *shape.get(2).unwrap_or(&1);
+        let data = self
+            .as_slice()
+            .ok_or(NdFirError)
+            .attach_printable("The ndarray is non continuous")?;
+        let data_bytes: &[u8] = bytemuck::cast_slice(data);
+
+        let pixel_type = to_pixel_type::<T>(channels)?;
+        ImageRef::new(cols, rows, data_bytes, pixel_type)
+            .change_context(NdFirError)
+            .attach_printable("Failed to create Image from ndarray")
+    }
+}
+
+impl<S: ndarray::DataMut<Elem = T>, T: seal::Sealed + bytemuck::Pod, D: ndarray::Dimension>
+    NdAsImageMut<T, D> for ndarray::ArrayBase<S, D>
+{
+    fn as_image_ref_mut(&mut self) -> Result<Image<'_>>
+    where
+        S: ndarray::DataMut<Elem = T>,
+    {
+        let shape = self.shape();
+        let rows = *shape
+            .first()
+            .ok_or_else(|| Report::new(NdFirError).attach_printable("Failed to get rows"))?
+            as u32;
+        let cols = *shape.get(1).unwrap_or(&1) as u32;
+        let channels = *shape.get(2).unwrap_or(&1);
+        let data = self
+            .as_slice_mut()
+            .ok_or(NdFirError)
+            .attach_printable("The ndarray is non continuous")?;
+        let data_bytes: &mut [u8] = bytemuck::cast_slice_mut(data);
+
+        let pixel_type = to_pixel_type::<T>(channels)?;
+        Image::from_slice_u8(cols, rows, data_bytes, pixel_type)
+            .change_context(NdFirError)
+            .attach_printable("Failed to create Image from ndarray")
+    }
+}
+
+pub trait NdFir<T, D> {
+    fn fast_resize<'o>(
+        &self,
+        height: usize,
+        width: usize,
+        options: impl Into<Option<&'o ResizeOptions>>,
+    ) -> Result<ndarray::Array<T, D>>;
+}
+
+impl<T: seal::Sealed + bytemuck::Pod + num::Zero, S: ndarray::Data<Elem = T>> NdFir<T, ndarray::Ix3>
+    for ndarray::ArrayBase<S, ndarray::Ix3>
+{
+    fn fast_resize<'o>(
+        &self,
+        height: usize,
+        width: usize,
+        options: impl Into<Option<&'o ResizeOptions>>,
+    ) -> Result<ndarray::Array3<T>> {
+        let source = self.as_image_ref()?;
+        let (_height, _width, channels) = self.dim();
+        let mut dest = ndarray::Array3::<T>::zeros((height, width, channels));
+        let mut dest_image = dest.as_image_ref_mut()?;
+        let mut resizer = fast_image_resize::Resizer::default();
+        resizer
+            .resize(&source, &mut dest_image, options)
+            .change_context(NdFirError)?;
+        Ok(dest)
+    }
+}
+
+impl<T: seal::Sealed + bytemuck::Pod + num::Zero, S: ndarray::Data<Elem = T>> NdFir<T, ndarray::Ix2>
+    for ndarray::ArrayBase<S, ndarray::Ix2>
+{
+    fn fast_resize<'o>(
+        &self,
+        height: usize,
+        width: usize,
+        options: impl Into<Option<&'o ResizeOptions>>,
+    ) -> Result<ndarray::Array<T, ndarray::Ix2>> {
+        let source = self.as_image_ref()?;
+        let (_height, _width) = self.dim();
+        let mut dest = ndarray::Array::<T, ndarray::Ix2>::zeros((height, width));
+        let mut dest_image = dest.as_image_ref_mut()?;
+        let mut resizer = fast_image_resize::Resizer::default();
+        resizer
+            .resize(&source, &mut dest_image, options)
+            .change_context(NdFirError)?;
+        Ok(dest)
+    }
+}
+
+#[test]
+pub fn test_ndarray_fast_image_resize_u8() {
+    let source_fhd = ndarray::Array3::<u8>::ones((1920, 1080, 3));
+    let mut resized_hd = ndarray::Array3::<u8>::zeros((1280, 720, 3));
+    let mut resizer = fast_image_resize::Resizer::default();
+    resizer
+        .resize(
+            &source_fhd.as_image_ref().unwrap(),
+            &mut resized_hd.as_image_ref_mut().unwrap(),
+            None,
+        )
+        .unwrap();
+    assert_eq!(resized_hd.shape(), [1280, 720, 3]);
+}

ndcv-bridge/src/gaussian.rs

@@ -0,0 +1,307 @@
+//! <https://docs.rs/opencv/latest/opencv/imgproc/fn.gaussian_blur.html>
+use crate::conversions::*;
+use crate::prelude_::*;
+use ndarray::*;
+
+#[repr(C)]
+#[derive(Default, Debug, Copy, Clone)]
+pub enum BorderType {
+    #[default]
+    BorderConstant = 0,
+    BorderReplicate = 1,
+    BorderReflect = 2,
+    BorderWrap = 3,
+    BorderReflect101 = 4,
+    BorderTransparent = 5,
+    BorderIsolated = 16,
+}
+
+#[repr(C)]
+#[derive(Default, Debug, Copy, Clone)]
+pub enum AlgorithmHint {
+    #[default]
+    AlgoHintDefault = 0,
+    AlgoHintAccurate = 1,
+    AlgoHintApprox = 2,
+}
+
+mod seal {
+    pub trait Sealed {}
+    // src: input image; the image can have any number of channels, which are processed independently, but the depth should be CV_8U, CV_16U, CV_16S, CV_32F or CV_64F.
+    impl Sealed for u8 {}
+    impl Sealed for u16 {}
+    impl Sealed for i16 {}
+    impl Sealed for f32 {}
+    impl Sealed for f64 {}
+}
+
+pub trait NdCvGaussianBlur<T: bytemuck::Pod + seal::Sealed, D: ndarray::Dimension>:
+    crate::image::NdImage + crate::conversions::NdAsImage<T, D>
+{
+    fn gaussian_blur(
+        &self,
+        kernel_size: (u8, u8),
+        sigma_x: f64,
+        sigma_y: f64,
+        border_type: BorderType,
+    ) -> Result<ndarray::Array<T, D>, NdCvError>;
+    fn gaussian_blur_def(
+        &self,
+        kernel: (u8, u8),
+        sigma_x: f64,
+    ) -> Result<ndarray::Array<T, D>, NdCvError> {
+        self.gaussian_blur(kernel, sigma_x, sigma_x, BorderType::BorderConstant)
+    }
+}
+
+impl<
+        T: bytemuck::Pod + num::Zero + seal::Sealed,
+        S: ndarray::RawData + ndarray::Data<Elem = T>,
+        D: ndarray::Dimension,
+    > NdCvGaussianBlur<T, D> for ArrayBase<S, D>
+where
+    ndarray::ArrayBase<S, D>: crate::image::NdImage + crate::conversions::NdAsImage<T, D>,
+    ndarray::Array<T, D>: crate::conversions::NdAsImageMut<T, D>,
+{
+    fn gaussian_blur(
+        &self,
+        kernel_size: (u8, u8),
+        sigma_x: f64,
+        sigma_y: f64,
+        border_type: BorderType,
+    ) -> Result<ndarray::Array<T, D>, NdCvError> {
+        let mut dst = ndarray::Array::zeros(self.dim());
+        let cv_self = self.as_image_mat()?;
+        let mut cv_dst = dst.as_image_mat_mut()?;
+        opencv::imgproc::gaussian_blur(
+            &*cv_self,
+            &mut *cv_dst,
+            opencv::core::Size {
+                width: kernel_size.0 as i32,
+                height: kernel_size.1 as i32,
+            },
+            sigma_x,
+            sigma_y,
+            border_type as i32,
+        )
+        .change_context(NdCvError)
+        .attach_printable("Failed to apply gaussian blur")?;
+        Ok(dst)
+    }
+}
+
+// impl<
+//         T: bytemuck::Pod + num::Zero + seal::Sealed,
+//         S: ndarray::RawData + ndarray::Data<Elem = T>,
+//     > NdCvGaussianBlur<T, Ix3> for ArrayBase<S, Ix3>
+// {
+//     fn gaussian_blur(
+//         &self,
+//         kernel_size: (u8, u8),
+//         sigma_x: f64,
+//         sigma_y: f64,
+//         border_type: BorderType,
+//     ) -> Result<ndarray::Array<T, Ix3>, NdCvError> {
+//         let mut dst = ndarray::Array::zeros(self.dim());
+//         let cv_self = self.as_image_mat()?;
+//         let mut cv_dst = dst.as_image_mat_mut()?;
+//         opencv::imgproc::gaussian_blur(
+//             &*cv_self,
+//             &mut *cv_dst,
+//             opencv::core::Size {
+//                 width: kernel_size.0 as i32,
+//                 height: kernel_size.1 as i32,
+//             },
+//             sigma_x,
+//             sigma_y,
+//             border_type as i32,
+//         )
+//         .change_context(NdCvError)
+//         .attach_printable("Failed to apply gaussian blur")?;
+//         Ok(dst)
+//     }
+// }
+//
+// impl<
+//         T: bytemuck::Pod + num::Zero + seal::Sealed,
+//         S: ndarray::RawData + ndarray::Data<Elem = T>,
+//     > NdCvGaussianBlur<T, Ix2> for ArrayBase<S, Ix2>
+// {
+//     fn gaussian_blur(
+//         &self,
+//         kernel_size: (u8, u8),
+//         sigma_x: f64,
+//         sigma_y: f64,
+//         border_type: BorderType,
+//     ) -> Result<ndarray::Array<T, Ix2>, NdCvError> {
+//         let mut dst = ndarray::Array::zeros(self.dim());
+//         let cv_self = self.as_image_mat()?;
+//         let mut cv_dst = dst.as_image_mat_mut()?;
+//         opencv::imgproc::gaussian_blur(
+//             &*cv_self,
+//             &mut *cv_dst,
+//             opencv::core::Size {
+//                 width: kernel_size.0 as i32,
+//                 height: kernel_size.1 as i32,
+//             },
+//             sigma_x,
+//             sigma_y,
+//             border_type as i32,
+//         )
+//         .change_context(NdCvError)
+//         .attach_printable("Failed to apply gaussian blur")?;
+//         Ok(dst)
+//     }
+// }
+
+/// For smaller values it is faster to use the allocated version
+/// For example in a 4k f32 image this is about 50% faster than the allocated one
+pub trait NdCvGaussianBlurInPlace<T: bytemuck::Pod + seal::Sealed, D: ndarray::Dimension>:
+    crate::image::NdImage + crate::conversions::NdAsImageMut<T, D>
+{
+    fn gaussian_blur_inplace(
+        &mut self,
+        kernel_size: (u8, u8),
+        sigma_x: f64,
+        sigma_y: f64,
+        border_type: BorderType,
+    ) -> Result<&mut Self, NdCvError>;
+    fn gaussian_blur_def_inplace(
+        &mut self,
+        kernel: (u8, u8),
+        sigma_x: f64,
+    ) -> Result<&mut Self, NdCvError> {
+        self.gaussian_blur_inplace(kernel, sigma_x, sigma_x, BorderType::BorderConstant)
+    }
+}
+
+impl<
+        T: bytemuck::Pod + num::Zero + seal::Sealed,
+        S: ndarray::RawData + ndarray::DataMut<Elem = T>,
+        D: ndarray::Dimension,
+    > NdCvGaussianBlurInPlace<T, D> for ArrayBase<S, D>
+where
+    Self: crate::image::NdImage + crate::conversions::NdAsImageMut<T, D>,
+{
+    fn gaussian_blur_inplace(
+        &mut self,
+        kernel_size: (u8, u8),
+        sigma_x: f64,
+        sigma_y: f64,
+        border_type: BorderType,
+    ) -> Result<&mut Self, NdCvError> {
+        let mut cv_self = self.as_image_mat_mut()?;
+
+        unsafe {
+            crate::inplace::op_inplace(&mut *cv_self, |this, out| {
+                opencv::imgproc::gaussian_blur(
+                    this,
+                    out,
+                    opencv::core::Size {
+                        width: kernel_size.0 as i32,
+                        height: kernel_size.1 as i32,
+                    },
+                    sigma_x,
+                    sigma_y,
+                    border_type as i32,
+                )
+            })
+        }
+        .change_context(NdCvError)
+        .attach_printable("Failed to apply gaussian blur")?;
+        Ok(self)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use ndarray::Array3;
+
+    #[test]
+    fn test_gaussian_basic() {
+        let arr = Array3::<u8>::ones((10, 10, 3));
+        let kernel_size = (3, 3);
+        let sigma_x = 0.0;
+        let sigma_y = 0.0;
+        let border_type = BorderType::BorderConstant;
+        let res = arr
+            .gaussian_blur(kernel_size, sigma_x, sigma_y, border_type)
+            .unwrap();
+        assert_eq!(res.shape(), &[10, 10, 3]);
+    }
+
+    #[test]
+    fn test_gaussian_edge_preservation() {
+        // Create an image with a sharp edge
+        let mut arr = Array3::<u8>::zeros((10, 10, 3));
+        arr.slice_mut(s![..5, .., ..]).fill(255); // Top half white, bottom half black
+
+        let res = arr
+            .gaussian_blur((3, 3), 1.0, 1.0, BorderType::BorderConstant)
+            .unwrap();
+
+        // Check that the middle row (edge) has intermediate values
+        let middle_row = res.slice(s![4..6, 5, 0]);
+        assert!(middle_row.iter().all(|&x| x > 0 && x < 255));
+    }
+
+    #[test]
+    fn test_gaussian_different_kernel_sizes() {
+        let arr = Array3::<u8>::ones((20, 20, 3));
+
+        // Test different kernel sizes
+        let kernel_sizes = [(3, 3), (5, 5), (7, 7)];
+        for &kernel_size in &kernel_sizes {
+            let res = arr
+                .gaussian_blur(kernel_size, 1.0, 1.0, BorderType::BorderConstant)
+                .unwrap();
+            assert_eq!(res.shape(), &[20, 20, 3]);
+        }
+    }
+
+    #[test]
+    fn test_gaussian_different_border_types() {
+        let mut arr = Array3::<u8>::zeros((10, 10, 3));
+        arr.slice_mut(s![4..7, 4..7, ..]).fill(255); // White square in center
+
+        let border_types = [
+            BorderType::BorderConstant,
+            BorderType::BorderReplicate,
+            BorderType::BorderReflect,
+            BorderType::BorderReflect101,
+        ];
+
+        for border_type in border_types {
+            let res = arr.gaussian_blur((3, 3), 1.0, 1.0, border_type).unwrap();
+            assert_eq!(res.shape(), &[10, 10, 3]);
+        }
+    }
+
+    #[test]
+    fn test_gaussian_different_types() {
+        // Test with different numeric types
+        let arr_u8 = Array3::<u8>::ones((10, 10, 3));
+        let arr_f32 = Array3::<f32>::ones((10, 10, 3));
+
+        let res_u8 = arr_u8
+            .gaussian_blur((3, 3), 1.0, 1.0, BorderType::BorderConstant)
+            .unwrap();
+        let res_f32 = arr_f32
+            .gaussian_blur((3, 3), 1.0, 1.0, BorderType::BorderConstant)
+            .unwrap();
+
+        assert_eq!(res_u8.shape(), &[10, 10, 3]);
+        assert_eq!(res_f32.shape(), &[10, 10, 3]);
+    }
+
+    #[test]
+    #[should_panic]
+    fn test_gaussian_invalid_kernel_size() {
+        let arr = Array3::<u8>::ones((10, 10, 3));
+        // Even kernel sizes should fail
+        let _ = arr
+            .gaussian_blur((2, 2), 1.0, 1.0, BorderType::BorderConstant)
+            .unwrap();
+    }
+}

ndcv-bridge/src/image.rs

@@ -0,0 +1,30 @@
+use ndarray::*;
+pub trait NdImage {
+    fn width(&self) -> usize;
+    fn height(&self) -> usize;
+    fn channels(&self) -> usize;
+}
+
+impl<T, S: RawData<Elem = T>> NdImage for ArrayBase<S, Ix3> {
+    fn width(&self) -> usize {
+        self.dim().1
+    }
+    fn height(&self) -> usize {
+        self.dim().0
+    }
+    fn channels(&self) -> usize {
+        self.dim().2
+    }
+}
+
+impl<T, S: RawData<Elem = T>> NdImage for ArrayBase<S, Ix2> {
+    fn width(&self) -> usize {
+        self.dim().1
+    }
+    fn height(&self) -> usize {
+        self.dim().0
+    }
+    fn channels(&self) -> usize {
+        1
+    }
+}

ndcv-bridge/src/inplace.rs

@@ -0,0 +1,14 @@
+use opencv::core::Mat;
+use opencv::prelude::*;
+use opencv::Result;
+
+#[inline(always)]
+pub(crate) unsafe fn op_inplace<T>(
+    m: &mut Mat,
+    f: impl FnOnce(&Mat, &mut Mat) -> Result<T>,
+) -> Result<T> {
+    let mut m_alias = Mat::from_raw(m.as_raw_mut());
+    let out = f(m, &mut m_alias);
+    let _ = m_alias.into_raw();
+    out
+}

ndcv-bridge/src/lib.rs

@@ -0,0 +1,83 @@
+//! Methods and type conversions for ndarray to opencv and vice versa
+mod blend;
+// mod dilate;
+pub mod fir;
+mod image;
+mod inplace;
+pub mod percentile;
+mod roi;
+
+#[cfg(feature = "opencv")]
+pub mod bounding_rect;
+// #[cfg(feature = "opencv")]
+// pub mod color_space;
+#[cfg(feature = "opencv")]
+pub mod connected_components;
+#[cfg(feature = "opencv")]
+pub mod contours;
+#[cfg(feature = "opencv")]
+pub mod conversions;
+// #[cfg(feature = "opencv")]
+// pub mod gaussian;
+#[cfg(feature = "opencv")]
+pub mod resize;
+
+pub mod codec;
+pub mod orient;
+pub use blend::NdBlend;
+pub use fast_image_resize::{FilterType, ResizeAlg, ResizeOptions, Resizer};
+pub use fir::NdFir;
+// pub use gaussian::{BorderType, NdCvGaussianBlur, NdCvGaussianBlurInPlace};
+pub use roi::{NdRoi, NdRoiMut, NdRoiZeroPadded};
+
+#[cfg(feature = "opencv")]
+pub use contours::{
+    ContourApproximationMethod, ContourHierarchy, ContourResult, ContourRetrievalMode,
+    NdCvContourArea, NdCvFindContours,
+};
+
+#[cfg(feature = "opencv")]
+pub use bounding_rect::BoundingRect;
+#[cfg(feature = "opencv")]
+pub use connected_components::{Connectivity, NdCvConnectedComponents};
+#[cfg(feature = "opencv")]
+pub use conversions::{MatAsNd, NdAsImage, NdAsImageMut, NdAsMat, NdAsMatMut, NdCvConversion};
+#[cfg(feature = "opencv")]
+pub use resize::{Interpolation, NdCvResize};
+
+pub(crate) mod prelude_ {
+    pub use crate::NdCvError;
+    pub use error_stack::*;
+}
+
+#[derive(Debug, thiserror::Error)]
+#[error("NdCvError")]
+pub struct NdCvError;
+
+#[cfg(feature = "opencv")]
+pub fn type_depth<T>() -> i32 {
+    match std::any::type_name::<T>() {
+        "u8" => opencv::core::CV_8U,
+        "i8" => opencv::core::CV_8S,
+        "u16" => opencv::core::CV_16U,
+        "i16" => opencv::core::CV_16S,
+        "i32" => opencv::core::CV_32S,
+        "f32" => opencv::core::CV_32F,
+        "f64" => opencv::core::CV_64F,
+        _ => panic!("Unsupported type"),
+    }
+}
+
+#[cfg(feature = "opencv")]
+pub fn depth_type(depth: i32) -> &'static str {
+    match depth {
+        opencv::core::CV_8U => "u8",
+        opencv::core::CV_8S => "i8",
+        opencv::core::CV_16U => "u16",
+        opencv::core::CV_16S => "i16",
+        opencv::core::CV_32S => "i32",
+        opencv::core::CV_32F => "f32",
+        opencv::core::CV_64F => "f64",
+        _ => panic!("Unsupported depth"),
+    }
+}

ndcv-bridge/src/orient.rs

@@ -0,0 +1,188 @@
+use ndarray::{Array, ArrayBase, ArrayView};
+
+#[derive(Clone, Copy)]
+pub enum Orientation {
+    NoRotation,
+    Mirror,
+    Clock180,
+    Water,
+    MirrorClock270,
+    Clock90,
+    MirrorClock90,
+    Clock270,
+    Unknown,
+}
+
+impl Orientation {
+    pub fn inverse(&self) -> Self {
+        match self {
+            Self::Clock90 => Self::Clock270,
+            Self::Clock270 => Self::Clock90,
+            _ => *self,
+        }
+    }
+}
+
+impl Orientation {
+    pub fn from_raw(flip: u8) -> Self {
+        match flip {
+            1 => Orientation::NoRotation,
+            2 => Orientation::Mirror,
+            3 => Orientation::Clock180,
+            4 => Orientation::Water,
+            5 => Orientation::MirrorClock270,
+            6 => Orientation::Clock90,
+            7 => Orientation::MirrorClock90,
+            8 => Orientation::Clock270,
+            _ => Orientation::Unknown,
+        }
+    }
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub enum RotationFlag {
+    Clock90,
+    Clock180,
+    Clock270,
+}
+
+impl RotationFlag {
+    pub fn neg(&self) -> Self {
+        match self {
+            RotationFlag::Clock90 => RotationFlag::Clock270,
+            RotationFlag::Clock180 => RotationFlag::Clock180,
+            RotationFlag::Clock270 => RotationFlag::Clock90,
+        }
+    }
+
+    pub fn to_orientation(&self) -> Orientation {
+        match self {
+            RotationFlag::Clock90 => Orientation::Clock90,
+            RotationFlag::Clock180 => Orientation::Clock180,
+            RotationFlag::Clock270 => Orientation::Clock270,
+        }
+    }
+}
+
+#[derive(Clone, Copy)]
+pub enum FlipFlag {
+    Mirror,
+    Water,
+    Both,
+}
+
+pub trait Orient<T: bytemuck::Pod, D: ndarray::Dimension> {
+    fn flip(&self, flip: FlipFlag) -> Array<T, D>;
+    fn rotate(&self, rotation: RotationFlag) -> Array<T, D>;
+    fn owned(&self) -> Array<T, D>;
+
+    fn unorient(&self, orientation: Orientation) -> Array<T, D>
+    where
+        Array<T, D>: Orient<T, D>,
+        Self: ToOwned<Owned = Array<T, D>>,
+    {
+        let inverse_orientation = orientation.inverse();
+        self.orient(inverse_orientation)
+
+        // match orientation {
+        //     Orientation::NoRotation | Orientation::Unknown => self.to_owned(),
+        //     Orientation::Mirror => self.flip(FlipFlag::Mirror).to_owned(),
+        //     Orientation::Clock180 => self.rotate(RotationFlag::Clock180),
+        //     Orientation::Water => self.flip(FlipFlag::Water).to_owned(),
+        //     Orientation::MirrorClock270 => self
+        //         .rotate(RotationFlag::Clock90)
+        //         .flip(FlipFlag::Mirror)
+        //         .to_owned(),
+        //     Orientation::Clock90 => self.rotate(RotationFlag::Clock270),
+        //     Orientation::MirrorClock90 => self
+        //         .rotate(RotationFlag::Clock270)
+        //         .flip(FlipFlag::Mirror)
+        //         .to_owned(),
+        //     Orientation::Clock270 => self.rotate(RotationFlag::Clock90),
+        // }
+    }
+
+    fn orient(&self, orientation: Orientation) -> Array<T, D>
+    where
+        Array<T, D>: Orient<T, D>,
+    {
+        match orientation {
+            Orientation::NoRotation | Orientation::Unknown => self.owned(),
+            Orientation::Mirror => self.flip(FlipFlag::Mirror).to_owned(),
+            Orientation::Clock180 => self.rotate(RotationFlag::Clock180),
+            Orientation::Water => self.flip(FlipFlag::Water).to_owned(),
+            Orientation::MirrorClock270 => self
+                .flip(FlipFlag::Mirror)
+                .rotate(RotationFlag::Clock270)
+                .to_owned(),
+            Orientation::Clock90 => self.rotate(RotationFlag::Clock90),
+            Orientation::MirrorClock90 => self
+                .flip(FlipFlag::Mirror)
+                .rotate(RotationFlag::Clock90)
+                .to_owned(),
+            Orientation::Clock270 => self.rotate(RotationFlag::Clock270),
+        }
+        .as_standard_layout()
+        .to_owned()
+    }
+}
+
+impl<T: bytemuck::Pod + Copy, S: ndarray::Data<Elem = T>> Orient<T, ndarray::Ix3>
+    for ArrayBase<S, ndarray::Ix3>
+{
+    fn flip(&self, flip: FlipFlag) -> Array<T, ndarray::Ix3> {
+        match flip {
+            FlipFlag::Mirror => self.slice(ndarray::s![.., ..;-1, ..]),
+            FlipFlag::Water => self.slice(ndarray::s![..;-1, .., ..]),
+            FlipFlag::Both => self.slice(ndarray::s![..;-1, ..;-1, ..]),
+        }
+        .as_standard_layout()
+        .to_owned()
+    }
+
+    fn owned(&self) -> Array<T, ndarray::Ix3> {
+        self.to_owned()
+    }
+
+    fn rotate(&self, rotation: RotationFlag) -> Array<T, ndarray::Ix3> {
+        match rotation {
+            RotationFlag::Clock90 => self
+                .view()
+                .permuted_axes([1, 0, 2])
+                .flip(FlipFlag::Mirror)
+                .to_owned(),
+            RotationFlag::Clock180 => self.flip(FlipFlag::Both).to_owned(),
+            RotationFlag::Clock270 => self
+                .view()
+                .permuted_axes([1, 0, 2])
+                .flip(FlipFlag::Water)
+                .to_owned(),
+        }
+    }
+}
+
+impl<T: bytemuck::Pod + Copy, S: ndarray::Data<Elem = T>> Orient<T, ndarray::Ix2>
+    for ArrayBase<S, ndarray::Ix2>
+{
+    fn flip(&self, flip: FlipFlag) -> Array<T, ndarray::Ix2> {
+        match flip {
+            FlipFlag::Mirror => self.slice(ndarray::s![.., ..;-1,]),
+            FlipFlag::Water => self.slice(ndarray::s![..;-1, ..,]),
+            FlipFlag::Both => self.slice(ndarray::s![..;-1, ..;-1,]),
+        }
+        .as_standard_layout()
+        .to_owned()
+    }
+
+    fn owned(&self) -> Array<T, ndarray::Ix2> {
+        self.to_owned()
+    }
+
+    fn rotate(&self, rotation: RotationFlag) -> Array<T, ndarray::Ix2> {
+        match rotation {
+            RotationFlag::Clock90 => self.t().flip(FlipFlag::Mirror).to_owned(),
+            RotationFlag::Clock180 => self.flip(FlipFlag::Both).to_owned(),
+            RotationFlag::Clock270 => self.t().flip(FlipFlag::Water).to_owned(),
+        }
+    }
+}

ndcv-bridge/src/percentile.rs

@@ -0,0 +1,63 @@
+use error_stack::*;
+use ndarray::{ArrayBase, Ix1};
+use num::cast::AsPrimitive;
+
+use crate::NdCvError;
+
+pub trait Percentile {
+    fn percentile(&self, qth_percentile: f64) -> Result<f64, NdCvError>;
+}
+
+impl<T: std::cmp::Ord + Clone + AsPrimitive<f64>, S: ndarray::Data<Elem = T>> Percentile
+    for ArrayBase<S, Ix1>
+{
+    fn percentile(&self, qth_percentile: f64) -> Result<f64, NdCvError> {
+        if self.len() == 0 {
+            return Err(error_stack::Report::new(NdCvError).attach_printable("Empty Input"));
+        }
+
+        if !(0_f64..1_f64).contains(&qth_percentile) {
+            return Err(error_stack::Report::new(NdCvError)
+                .attach_printable("Qth percentile must be between 0 and 1"));
+        }
+
+        let mut standard_array = self.as_standard_layout();
+        let mut raw_data = standard_array
+            .as_slice_mut()
+            .expect("An array in standard layout will always return its inner slice");
+
+        raw_data.sort();
+
+        let actual_index = qth_percentile * (raw_data.len() - 1) as f64;
+
+        let lower_index = (actual_index.floor() as usize).clamp(0, raw_data.len() - 1);
+        let upper_index = (actual_index.ceil() as usize).clamp(0, raw_data.len() - 1);
+
+        if lower_index == upper_index {
+            Ok(raw_data[lower_index].as_())
+        } else {
+            let weight = actual_index - lower_index as f64;
+            Ok(raw_data[lower_index].as_() * (1.0 - weight) + raw_data[upper_index].as_() * weight)
+        }
+    }
+}
+
+// fn percentile(data: &Array1<f64>, p: f64) -> f64 {
+//     if data.len() == 0 {
+//         return 0.0;
+//     }
+//
+//     let mut sorted_data = data.to_vec();
+//     sorted_data.sort_by(|a, b| a.partial_cmp(b).unwrap());
+//
+//     let index = (p / 100.0) * (sorted_data.len() - 1) as f64;
+//     let lower = index.floor() as usize;
+//     let upper = index.ceil() as usize;
+//
+//     if lower == upper {
+//         sorted_data[lower] as f64
+//     } else {
+//         let weight = index - lower as f64;
+//         sorted_data[lower] as f64 * (1.0 - weight) + sorted_data[upper] as f64 * weight
+//     }
+// }

ndcv-bridge/src/resize.rs

@@ -0,0 +1,108 @@
+use crate::{prelude_::*, NdAsImage, NdAsImageMut};
+
+/// Resize ndarray using OpenCV resize functions
+pub trait NdCvResize<T, D>: seal::SealedInternal {
+    /// The input array must be a continuous 2D or 3D ndarray
+    fn resize(
+        &self,
+        height: u16,
+        width: u16,
+        interpolation: Interpolation,
+    ) -> Result<ndarray::ArrayBase<ndarray::OwnedRepr<T>, D>, NdCvError>;
+}
+
+#[repr(i32)]
+#[derive(Debug, Copy, Clone)]
+pub enum Interpolation {
+    Linear = opencv::imgproc::INTER_LINEAR,
+    LinearExact = opencv::imgproc::INTER_LINEAR_EXACT,
+    Max = opencv::imgproc::INTER_MAX,
+    Area = opencv::imgproc::INTER_AREA,
+    Cubic = opencv::imgproc::INTER_CUBIC,
+    Nearest = opencv::imgproc::INTER_NEAREST,
+    NearestExact = opencv::imgproc::INTER_NEAREST_EXACT,
+    Lanczos4 = opencv::imgproc::INTER_LANCZOS4,
+}
+
+mod seal {
+    pub trait SealedInternal {}
+    impl<T: bytemuck::Pod, S: ndarray::Data<Elem = T>> SealedInternal
+        for ndarray::ArrayBase<S, ndarray::Ix3>
+    {
+    }
+    impl<T: bytemuck::Pod, S: ndarray::Data<Elem = T>> SealedInternal
+        for ndarray::ArrayBase<S, ndarray::Ix2>
+    {
+    }
+}
+
+impl<T: bytemuck::Pod + num::Zero, S: ndarray::Data<Elem = T>> NdCvResize<T, ndarray::Ix2>
+    for ndarray::ArrayBase<S, ndarray::Ix2>
+{
+    fn resize(
+        &self,
+        height: u16,
+        width: u16,
+        interpolation: Interpolation,
+    ) -> Result<ndarray::Array2<T>, NdCvError> {
+        let mat = self.as_image_mat()?;
+        let mut dest = ndarray::Array2::zeros((height.into(), width.into()));
+        let mut dest_mat = dest.as_image_mat_mut()?;
+        opencv::imgproc::resize(
+            mat.as_ref(),
+            dest_mat.as_mut(),
+            opencv::core::Size {
+                height: height.into(),
+                width: width.into(),
+            },
+            0.,
+            0.,
+            interpolation as i32,
+        )
+        .change_context(NdCvError)?;
+        Ok(dest)
+    }
+}
+
+impl<T: bytemuck::Pod + num::Zero, S: ndarray::Data<Elem = T>> NdCvResize<T, ndarray::Ix3>
+    for ndarray::ArrayBase<S, ndarray::Ix3>
+{
+    fn resize(
+        &self,
+        height: u16,
+        width: u16,
+        interpolation: Interpolation,
+    ) -> Result<ndarray::ArrayBase<ndarray::OwnedRepr<T>, ndarray::Ix3>, NdCvError> {
+        let mat = self.as_image_mat()?;
+        let mut dest =
+            ndarray::Array3::zeros((height.into(), width.into(), self.len_of(ndarray::Axis(2))));
+        let mut dest_mat = dest.as_image_mat_mut()?;
+        opencv::imgproc::resize(
+            mat.as_ref(),
+            dest_mat.as_mut(),
+            opencv::core::Size {
+                height: height.into(),
+                width: width.into(),
+            },
+            0.,
+            0.,
+            interpolation as i32,
+        )
+        .change_context(NdCvError)?;
+        Ok(dest)
+    }
+}
+
+#[test]
+fn test_resize_simple() {
+    let foo = ndarray::Array2::<u8>::ones((10, 10));
+    let foo_resized = foo.resize(15, 20, Interpolation::Linear).unwrap();
+    assert_eq!(foo_resized, ndarray::Array2::<u8>::ones((15, 20)));
+}
+
+#[test]
+fn test_resize_3d() {
+    let foo = ndarray::Array3::<u8>::ones((10, 10, 3));
+    let foo_resized = foo.resize(15, 20, Interpolation::Linear).unwrap();
+    assert_eq!(foo_resized, ndarray::Array3::<u8>::ones((15, 20, 3)));
+}

ndcv-bridge/src/roi.rs

@@ -0,0 +1,274 @@
+pub trait NdRoi<T, D>: seal::Sealed {
+    fn roi(&self, rect: bounding_box::Aabb2<usize>) -> ndarray::ArrayView<T, D>;
+}
+
+pub trait NdRoiMut<T, D>: seal::Sealed {
+    fn roi_mut(&mut self, rect: bounding_box::Aabb2<usize>) -> ndarray::ArrayViewMut<T, D>;
+}
+
+mod seal {
+    use ndarray::{Ix2, Ix3};
+    pub trait Sealed {}
+    impl<T: bytemuck::Pod, S: ndarray::Data<Elem = T>> Sealed for ndarray::ArrayBase<S, Ix2> {}
+    impl<T: bytemuck::Pod, S: ndarray::Data<Elem = T>> Sealed for ndarray::ArrayBase<S, Ix3> {}
+}
+
+impl<T: bytemuck::Pod, S: ndarray::Data<Elem = T>> NdRoi<T, ndarray::Ix3>
+    for ndarray::ArrayBase<S, ndarray::Ix3>
+{
+    fn roi(&self, rect: bounding_box::Aabb2<usize>) -> ndarray::ArrayView3<T> {
+        let y1 = rect.y1();
+        let y2 = rect.y2();
+        let x1 = rect.x1();
+        let x2 = rect.x2();
+        self.slice(ndarray::s![y1..y2, x1..x2, ..])
+    }
+}
+
+impl<T: bytemuck::Pod, S: ndarray::DataMut<Elem = T>> NdRoiMut<T, ndarray::Ix3>
+    for ndarray::ArrayBase<S, ndarray::Ix3>
+{
+    fn roi_mut(&mut self, rect: bounding_box::Aabb2<usize>) -> ndarray::ArrayViewMut3<T> {
+        let y1 = rect.y1();
+        let y2 = rect.y2();
+        let x1 = rect.x1();
+        let x2 = rect.x2();
+        self.slice_mut(ndarray::s![y1..y2, x1..x2, ..])
+    }
+}
+
+impl<T: bytemuck::Pod, S: ndarray::Data<Elem = T>> NdRoi<T, ndarray::Ix2>
+    for ndarray::ArrayBase<S, ndarray::Ix2>
+{
+    fn roi(&self, rect: bounding_box::Aabb2<usize>) -> ndarray::ArrayView2<T> {
+        let y1 = rect.y1();
+        let y2 = rect.y2();
+        let x1 = rect.x1();
+        let x2 = rect.x2();
+        self.slice(ndarray::s![y1..y2, x1..x2])
+    }
+}
+
+impl<T: bytemuck::Pod, S: ndarray::DataMut<Elem = T>> NdRoiMut<T, ndarray::Ix2>
+    for ndarray::ArrayBase<S, ndarray::Ix2>
+{
+    fn roi_mut(&mut self, rect: bounding_box::Aabb2<usize>) -> ndarray::ArrayViewMut2<T> {
+        let y1 = rect.y1();
+        let y2 = rect.y2();
+        let x1 = rect.x1();
+        let x2 = rect.x2();
+        self.slice_mut(ndarray::s![y1..y2, x1..x2])
+    }
+}
+
+#[test]
+fn test_roi() {
+    let arr = ndarray::Array3::<u8>::zeros((10, 10, 3));
+    let rect = bounding_box::Aabb2::from_xywh(1, 1, 3, 3);
+    let roi = arr.roi(rect);
+    assert_eq!(roi.shape(), &[3, 3, 3]);
+}
+
+#[test]
+fn test_roi_2d() {
+    let arr = ndarray::Array2::<u8>::zeros((10, 10));
+    let rect = bounding_box::Aabb2::from_xywh(1, 1, 3, 3);
+    let roi = arr.roi(rect);
+    assert_eq!(roi.shape(), &[3, 3]);
+}
+
+/// ```text
+///    ┌──────────────────┐
+///    │      padded      │
+///    │    ┌────────┐    │
+///    │    │        │    │
+///    │    │original│    │
+///    │    │        │    │
+///    │    └────────┘    │
+///    │      zeroed      │
+///    └──────────────────┘
+/// ```
+///
+/// Returns the padded bounding box and the padded segment
+/// The padded is the padded bounding box
+/// The original is the original bounding box
+/// Returns the padded bounding box as zeros and the original bbox as the original segment
+// Helper functions for missing methods from old bbox crate
+fn bbox_top_left_usize(bbox: &bounding_box::Aabb2<usize>) -> (usize, usize) {
+    (bbox.x1(), bbox.y1())
+}
+
+fn bbox_with_top_left_usize(
+    bbox: &bounding_box::Aabb2<usize>,
+    x: usize,
+    y: usize,
+) -> bounding_box::Aabb2<usize> {
+    let width = bbox.x2() - bbox.x1();
+    let height = bbox.y2() - bbox.y1();
+    bounding_box::Aabb2::from_xywh(x, y, width, height)
+}
+
+fn bbox_with_origin_usize(point: (usize, usize), origin: (usize, usize)) -> (usize, usize) {
+    (point.0 - origin.0, point.1 - origin.1)
+}
+
+fn bbox_zeros_ndarray_2d<T: num::Zero + Copy>(
+    bbox: &bounding_box::Aabb2<usize>,
+) -> ndarray::Array2<T> {
+    let width = bbox.x2() - bbox.x1();
+    let height = bbox.y2() - bbox.y1();
+    ndarray::Array2::<T>::zeros((height, width))
+}
+
+fn bbox_zeros_ndarray_3d<T: num::Zero + Copy>(
+    bbox: &bounding_box::Aabb2<usize>,
+    channels: usize,
+) -> ndarray::Array3<T> {
+    let width = bbox.x2() - bbox.x1();
+    let height = bbox.y2() - bbox.y1();
+    ndarray::Array3::<T>::zeros((height, width, channels))
+}
+
+fn bbox_round_f64(bbox: &bounding_box::Aabb2<f64>) -> bounding_box::Aabb2<f64> {
+    let x1 = bbox.x1().round();
+    let y1 = bbox.y1().round();
+    let x2 = bbox.x2().round();
+    let y2 = bbox.y2().round();
+    bounding_box::Aabb2::from_x1y1x2y2(x1, y1, x2, y2)
+}
+
+fn bbox_cast_f64_to_usize(bbox: &bounding_box::Aabb2<f64>) -> bounding_box::Aabb2<usize> {
+    let x1 = bbox.x1() as usize;
+    let y1 = bbox.y1() as usize;
+    let x2 = bbox.x2() as usize;
+    let y2 = bbox.y2() as usize;
+    bounding_box::Aabb2::from_x1y1x2y2(x1, y1, x2, y2)
+}
+
+pub trait NdRoiZeroPadded<T, D: ndarray::Dimension>: seal::Sealed {
+    fn roi_zero_padded(
+        &self,
+        original: bounding_box::Aabb2<usize>,
+        padded: bounding_box::Aabb2<usize>,
+    ) -> (bounding_box::Aabb2<usize>, ndarray::Array<T, D>);
+}
+
+impl<T: bytemuck::Pod + num::Zero> NdRoiZeroPadded<T, ndarray::Ix2> for ndarray::Array2<T> {
+    fn roi_zero_padded(
+        &self,
+        original: bounding_box::Aabb2<usize>,
+        padded: bounding_box::Aabb2<usize>,
+    ) -> (bounding_box::Aabb2<usize>, ndarray::Array2<T>) {
+        // The co-ordinates of both the original and the padded bounding boxes must be contained in
+        // self or it will panic
+
+        let self_bbox = bounding_box::Aabb2::from_xywh(0, 0, self.shape()[1], self.shape()[0]);
+        if !self_bbox.contains_bbox(&original) {
+            panic!("original bounding box is not contained in self");
+        }
+        if !self_bbox.contains_bbox(&padded) {
+            panic!("padded bounding box is not contained in self");
+        }
+
+        let original_top_left = bbox_top_left_usize(&original);
+        let padded_top_left = bbox_top_left_usize(&padded);
+        let origin_offset = bbox_with_origin_usize(original_top_left, padded_top_left);
+        let original_roi_in_padded =
+            bbox_with_top_left_usize(&original, origin_offset.0, origin_offset.1);
+
+        let original_segment = self.roi(original);
+        let mut padded_segment = bbox_zeros_ndarray_2d::<T>(&padded);
+        padded_segment
+            .roi_mut(original_roi_in_padded)
+            .assign(&original_segment);
+
+        (padded, padded_segment)
+    }
+}
+
+impl<T: bytemuck::Pod + num::Zero> NdRoiZeroPadded<T, ndarray::Ix3> for ndarray::Array3<T> {
+    fn roi_zero_padded(
+        &self,
+        original: bounding_box::Aabb2<usize>,
+        padded: bounding_box::Aabb2<usize>,
+    ) -> (bounding_box::Aabb2<usize>, ndarray::Array3<T>) {
+        let self_bbox = bounding_box::Aabb2::from_xywh(0, 0, self.shape()[1], self.shape()[0]);
+        if !self_bbox.contains_bbox(&original) {
+            panic!("original bounding box is not contained in self");
+        }
+        if !self_bbox.contains_bbox(&padded) {
+            panic!("padded bounding box is not contained in self");
+        }
+
+        let original_top_left = bbox_top_left_usize(&original);
+        let padded_top_left = bbox_top_left_usize(&padded);
+        let origin_offset = bbox_with_origin_usize(original_top_left, padded_top_left);
+        let original_roi_in_padded =
+            bbox_with_top_left_usize(&original, origin_offset.0, origin_offset.1);
+
+        let original_segment = self.roi(original);
+        let mut padded_segment = bbox_zeros_ndarray_3d::<T>(&padded, self.len_of(ndarray::Axis(2)));
+        padded_segment
+            .roi_mut(original_roi_in_padded)
+            .assign(&original_segment);
+
+        (padded, padded_segment)
+    }
+}
+
+#[test]
+fn test_roi_zero_padded() {
+    let arr = ndarray::Array2::<u8>::ones((10, 10));
+    let original = bounding_box::Aabb2::from_xywh(1.0, 1.0, 3.0, 3.0);
+    let clamp = bounding_box::Aabb2::from_xywh(0.0, 0.0, 10.0, 10.0);
+    let padded = original.padding(2.0).clamp(&clamp).unwrap();
+    let padded_cast = bbox_cast_f64_to_usize(&padded);
+    let original_cast = bbox_cast_f64_to_usize(&original);
+    let (padded_result, padded_segment) = arr.roi_zero_padded(original_cast, padded_cast);
+    assert_eq!(padded_result, bounding_box::Aabb2::from_xywh(0, 0, 6, 6));
+    assert_eq!(padded_segment.shape(), &[6, 6]);
+}
+
+#[test]
+pub fn bbox_clamp_failure_preload() {
+    let segment_mask = ndarray::Array2::<u8>::zeros((512, 512));
+    let og = bounding_box::Aabb2::from_xywh(475.0, 79.625, 37.0, 282.15);
+    let clamp = bounding_box::Aabb2::from_xywh(0.0, 0.0, 512.0, 512.0);
+    let padded = og
+        .scale(nalgebra::Vector2::new(1.2, 1.2))
+        .clamp(&clamp)
+        .unwrap();
+    let padded = bbox_round_f64(&padded);
+    let og_cast = bbox_cast_f64_to_usize(&bbox_round_f64(&og));
+    let padded_cast = bbox_cast_f64_to_usize(&padded);
+    let (_bbox, _segment) = segment_mask.roi_zero_padded(og_cast, padded_cast);
+}
+
+#[test]
+pub fn bbox_clamp_failure_preload_2() {
+    let segment_mask = ndarray::Array2::<u8>::zeros((512, 512));
+    let bbox = bounding_box::Aabb2::from_xywh(354.25, 98.5, 116.25, 413.5);
+    // let padded = bounding_box::Aabb2::from_xywh(342.625, 57.150000000000006, 139.5, 454.85);
+    let clamp = bounding_box::Aabb2::from_xywh(0.0, 0.0, 512.0, 512.0);
+    let padded = bbox
+        .scale(nalgebra::Vector2::new(1.2, 1.2))
+        .clamp(&clamp)
+        .unwrap();
+    let padded = bbox_round_f64(&padded);
+    let bbox_cast = bbox_cast_f64_to_usize(&bbox_round_f64(&bbox));
+    let padded_cast = bbox_cast_f64_to_usize(&padded);
+    let (_bbox, _segment) = segment_mask.roi_zero_padded(bbox_cast, padded_cast);
+}
+
+#[test]
+fn test_roi_zero_padded_3d() {
+    let arr = ndarray::Array3::<u8>::ones((10, 10, 3));
+    let original = bounding_box::Aabb2::from_xywh(1.0, 1.0, 3.0, 3.0);
+    let clamp = bounding_box::Aabb2::from_xywh(0.0, 0.0, 10.0, 10.0);
+    let padded = original.padding(2.0).clamp(&clamp).unwrap();
+    let padded_cast = bbox_cast_f64_to_usize(&padded);
+    let original_cast = bbox_cast_f64_to_usize(&original);
+    let (padded_result, padded_segment) = arr.roi_zero_padded(original_cast, padded_cast);
+    assert_eq!(padded_result, bounding_box::Aabb2::from_xywh(0, 0, 6, 6));
+    assert_eq!(padded_segment.shape(), &[6, 6, 3]);
+}

patches/ort_env_global_mutex.patch

@@ -0,0 +1,42 @@
+From 83e1dbf52b7695a2966795e0350aaa385d1ba8c8 Mon Sep 17 00:00:00 2001
+From: "Carson M." <carson@pyke.io>
+Date: Sun, 22 Jun 2025 23:53:20 -0500
+Subject: [PATCH] Leak logger mutex
+
+---
+ onnxruntime/core/common/logging/logging.cc | 8 ++++----
+ 1 file changed, 4 insertions(+), 4 deletions(-)
+
+diff --git a/onnxruntime/core/common/logging/logging.cc b/onnxruntime/core/common/logging/logging.cc
+index a79e7300cffce..07578fc72ca99 100644
+--- a/onnxruntime/core/common/logging/logging.cc
++++ b/onnxruntime/core/common/logging/logging.cc
+@@ -64,8 +64,8 @@ LoggingManager* LoggingManager::GetDefaultInstance() {
+ #pragma warning(disable : 26426)
+ #endif
+ 
+-static std::mutex& DefaultLoggerMutex() noexcept {
+-  static std::mutex mutex;
++static std::mutex* DefaultLoggerMutex() noexcept {
++  static std::mutex* mutex = new std::mutex();
+   return mutex;
+ }
+ 
+@@ -107,7 +107,7 @@ LoggingManager::LoggingManager(std::unique_ptr<ISink> sink, Severity default_min
+ 
+     // lock mutex to create instance, and enable logging
+     // this matches the mutex usage in Shutdown
+-    std::lock_guard<std::mutex> guard(DefaultLoggerMutex());
++    std::lock_guard<std::mutex> guard(*DefaultLoggerMutex());
+ 
+     if (DefaultLoggerManagerInstance().load() != nullptr) {
+       ORT_THROW("Only one instance of LoggingManager created with InstanceType::Default can exist at any point in time.");
+@@ -127,7 +127,7 @@ LoggingManager::LoggingManager(std::unique_ptr<ISink> sink, Severity default_min
+ LoggingManager::~LoggingManager() {
+   if (owns_default_logger_) {
+     // lock mutex to reset DefaultLoggerManagerInstance() and free default logger from this instance.
+-    std::lock_guard<std::mutex> guard(DefaultLoggerMutex());
++    std::lock_guard<std::mutex> guard(*DefaultLoggerMutex());
+ #if ((__cplusplus >= 201703L) || (defined(_MSVC_LANG) && (_MSVC_LANG >= 201703L)))
+     DefaultLoggerManagerInstance().store(nullptr, std::memory_order_release);
+ #else

sqlite3-ndarray-math/Cargo.toml

@@ -0,0 +1,14 @@
+[package]
+name = "sqlite3-ndarray-math"
+version.workspace = true
+edition.workspace = true
+
+[lib]
+crate-type = ["cdylib", "staticlib"]
+
+[dependencies]
+ndarray = "0.16.1"
+ndarray-math = { git = "https://git.darksailor.dev/servius/ndarray-math", version = "0.1.0" }
+# ndarray-math = { path = "/Users/fs0c131y/Projects/ndarray-math", version = "0.1.0" }
+ndarray-safetensors = { version = "0.1.0", path = "../ndarray-safetensors" }
+sqlite-loadable = "0.0.5"

sqlite3-ndarray-math/src/lib.rs

@@ -0,0 +1,61 @@
+use sqlite_loadable::prelude::*;
+use sqlite_loadable::{Error, ErrorKind};
+use sqlite_loadable::{Result, api, define_scalar_function};
+
+fn cosine_similarity(context: *mut sqlite3_context, values: &[*mut sqlite3_value]) -> Result<()> {
+    #[inline(always)]
+    fn custom_error(err: impl core::error::Error) -> sqlite_loadable::Error {
+        sqlite_loadable::Error::new(sqlite_loadable::ErrorKind::Message(err.to_string()))
+    }
+
+    if values.len() != 2 {
+        return Err(Error::new(ErrorKind::Message(
+            "cosine_similarity requires exactly 2 arguments".to_string(),
+        )));
+    }
+    let array_1 = api::value_blob(values.get(0).expect("1st argument"));
+    let array_2 = api::value_blob(values.get(1).expect("2nd argument"));
+    let array_1_st =
+        ndarray_safetensors::SafeArraysView::from_bytes(array_1).map_err(custom_error)?;
+    let array_2_st =
+        ndarray_safetensors::SafeArraysView::from_bytes(array_2).map_err(custom_error)?;
+
+    let array_view_1 = array_1_st
+        .tensor_by_index::<f32, ndarray::Ix1>(0)
+        .map_err(custom_error)?;
+    let array_view_2 = array_2_st
+        .tensor_by_index::<f32, ndarray::Ix1>(0)
+        .map_err(custom_error)?;
+
+    use ndarray_math::*;
+    let similarity = array_view_1
+        .cosine_similarity(array_view_2)
+        .map_err(custom_error)?;
+    api::result_double(context, similarity as f64);
+
+    Ok(())
+}
+
+pub fn _sqlite3_extension_init(db: *mut sqlite3) -> Result<()> {
+    define_scalar_function(
+        db,
+        "cosine_similarity",
+        2,
+        cosine_similarity,
+        FunctionFlags::DETERMINISTIC,
+    )?;
+    Ok(())
+}
+
+/// # Safety
+///
+/// Should only be called by underlying SQLite C APIs,
+/// like sqlite3_auto_extension and sqlite3_cancel_auto_extension.
+#[unsafe(no_mangle)]
+pub unsafe extern "C" fn sqlite3_extension_init(
+    db: *mut sqlite3,
+    pz_err_msg: *mut *mut c_char,
+    p_api: *mut sqlite3_api_routines,
+) -> c_uint {
+    register_entrypoint(db, pz_err_msg, p_api, _sqlite3_extension_init)
+}

src/bin/detector-cli/cli.rs

@@ -0,0 +1,213 @@
+use detector::ort_ep;
+use std::path::PathBuf;
+
+#[derive(Debug, clap::Parser)]
+pub struct Cli {
+    #[clap(subcommand)]
+    pub cmd: SubCommand,
+}
+
+#[derive(Debug, clap::Subcommand)]
+pub enum SubCommand {
+    #[clap(name = "detect")]
+    Detect(Detect),
+    #[clap(name = "detect-multi")]
+    DetectMulti(DetectMulti),
+    #[clap(name = "query")]
+    Query(Query),
+    #[clap(name = "similar")]
+    Similar(Similar),
+    #[clap(name = "stats")]
+    Stats(Stats),
+    #[clap(name = "compare")]
+    Compare(Compare),
+    #[clap(name = "cluster")]
+    Cluster(Cluster),
+    #[clap(name = "gui")]
+    Gui,
+    #[clap(name = "completions")]
+    Completions { shell: clap_complete::Shell },
+}
+
+#[derive(Debug, clap::ValueEnum, Clone, Copy, PartialEq)]
+pub enum Models {
+    RetinaFace,
+    Yolo,
+}
+
+#[derive(Debug, Clone)]
+pub enum Executor {
+    Mnn(mnn::ForwardType),
+    Ort(Vec<ort_ep::ExecutionProvider>),
+}
+
+#[derive(Debug, clap::Args)]
+pub struct Detect {
+    #[clap(short, long)]
+    pub model: Option<PathBuf>,
+    #[clap(short = 'M', long, default_value = "retina-face")]
+    pub model_type: Models,
+    #[clap(short, long)]
+    pub output: Option<PathBuf>,
+    #[clap(
+        short = 'p',
+        long,
+        default_value = "cpu",
+        group = "execution_provider",
+        required_unless_present = "mnn_forward_type"
+    )]
+    pub ort_execution_provider: Vec<ort_ep::ExecutionProvider>,
+    #[clap(
+        short = 'f',
+        long,
+        group = "execution_provider",
+        required_unless_present = "ort_execution_provider"
+    )]
+    pub mnn_forward_type: Option<mnn::ForwardType>,
+    #[clap(short, long, default_value_t = 0.8)]
+    pub threshold: f32,
+    #[clap(short, long, default_value_t = 0.3)]
+    pub nms_threshold: f32,
+    #[clap(short, long, default_value_t = 8)]
+    pub batch_size: usize,
+    #[clap(short = 'd', long)]
+    pub database: Option<PathBuf>,
+    #[clap(long, default_value = "facenet")]
+    pub model_name: String,
+    #[clap(long)]
+    pub save_to_db: bool,
+    pub image: PathBuf,
+}
+
+#[derive(Debug, clap::Args)]
+pub struct DetectMulti {
+    #[clap(short, long)]
+    pub model: Option<PathBuf>,
+    #[clap(short = 'M', long, default_value = "retina-face")]
+    pub model_type: Models,
+    #[clap(short, long)]
+    pub output_dir: Option<PathBuf>,
+    #[clap(
+        short = 'p',
+        long,
+        default_value = "cpu",
+        group = "execution_provider",
+        required_unless_present = "mnn_forward_type"
+    )]
+    pub ort_execution_provider: Vec<ort_ep::ExecutionProvider>,
+    #[clap(
+        short = 'f',
+        long,
+        group = "execution_provider",
+        required_unless_present = "ort_execution_provider"
+    )]
+    pub mnn_forward_type: Option<mnn::ForwardType>,
+    #[clap(short, long, default_value_t = 0.8)]
+    pub threshold: f32,
+    #[clap(short, long, default_value_t = 0.3)]
+    pub nms_threshold: f32,
+    #[clap(short, long, default_value_t = 8)]
+    pub batch_size: usize,
+    #[clap(short = 'd', long, default_value = "face_detections.db")]
+    pub database: PathBuf,
+    #[clap(long, default_value = "facenet")]
+    pub model_name: String,
+    #[clap(
+        long,
+        help = "Image extensions to process (e.g., jpg,png,jpeg)",
+        default_value = "jpg,jpeg,png,bmp,tiff,webp"
+    )]
+    pub extensions: String,
+    #[clap(help = "Directory containing images to process")]
+    pub input_dir: PathBuf,
+}
+
+#[derive(Debug, clap::Args)]
+pub struct Query {
+    #[clap(short = 'd', long, default_value = "face_detections.db")]
+    pub database: PathBuf,
+    #[clap(short, long)]
+    pub image_id: Option<i64>,
+    #[clap(short, long)]
+    pub face_id: Option<i64>,
+    #[clap(long)]
+    pub show_embeddings: bool,
+    #[clap(long)]
+    pub show_landmarks: bool,
+}
+
+#[derive(Debug, clap::Args)]
+pub struct Similar {
+    #[clap(short = 'd', long, default_value = "face_detections.db")]
+    pub database: PathBuf,
+    #[clap(short, long)]
+    pub face_id: i64,
+    #[clap(short, long, default_value_t = 0.7)]
+    pub threshold: f32,
+    #[clap(short, long, default_value_t = 10)]
+    pub limit: usize,
+}
+
+#[derive(Debug, clap::Args)]
+pub struct Stats {
+    #[clap(short = 'd', long, default_value = "face_detections.db")]
+    pub database: PathBuf,
+}
+
+#[derive(Debug, clap::Args)]
+pub struct Compare {
+    #[clap(short, long)]
+    pub model: Option<PathBuf>,
+    #[clap(short = 'M', long, default_value = "retina-face")]
+    pub model_type: Models,
+    #[clap(
+        short = 'p',
+        long,
+        default_value = "cpu",
+        group = "execution_provider",
+        required_unless_present = "mnn_forward_type"
+    )]
+    pub ort_execution_provider: Vec<ort_ep::ExecutionProvider>,
+    #[clap(
+        short = 'f',
+        long,
+        group = "execution_provider",
+        required_unless_present = "ort_execution_provider"
+    )]
+    pub mnn_forward_type: Option<mnn::ForwardType>,
+    #[clap(short, long, default_value_t = 0.8)]
+    pub threshold: f32,
+    #[clap(short, long, default_value_t = 0.3)]
+    pub nms_threshold: f32,
+    #[clap(short, long, default_value_t = 8)]
+    pub batch_size: usize,
+    #[clap(long, default_value = "facenet")]
+    pub model_name: String,
+    #[clap(help = "First image to compare")]
+    pub image1: PathBuf,
+    #[clap(help = "Second image to compare")]
+    pub image2: PathBuf,
+}
+
+#[derive(Debug, clap::Args)]
+pub struct Cluster {
+    #[clap(short = 'd', long, default_value = "face_detections.db")]
+    pub database: PathBuf,
+    #[clap(short, long, default_value_t = 5)]
+    pub clusters: usize,
+    #[clap(short, long, default_value_t = 100)]
+    pub max_iterations: u64,
+    #[clap(short, long, default_value_t = 1e-4)]
+    pub tolerance: f64,
+    #[clap(long, default_value = "facenet")]
+    pub model_name: String,
+    #[clap(short, long)]
+    pub output: Option<PathBuf>,
+}
+
+impl Cli {
+    pub fn completions(shell: clap_complete::Shell) {
+        let mut command = <Cli as clap::CommandFactory>::command();
+        clap_complete::generate(shell, &mut command, "detector", &mut std::io::stdout());
+    }
+}

src/bin/gui.rs

@@ -0,0 +1,19 @@
+use detector::errors::*;
+fn main() -> Result<()> {
+    // Initialize logging
+    tracing_subscriber::fmt()
+        .with_env_filter("warn,ort=warn")
+        .with_file(true)
+        .with_line_number(true)
+        // .with_thread_names(true)
+        .with_target(true)
+        .init();
+
+    // Run the GUI
+    if let Err(e) = detector::gui::run() {
+        eprintln!("GUI error: {}", e);
+        std::process::exit(1);
+    }
+
+    Ok(())
+}

src/cli.rs

@@ -1,71 +0,0 @@
-use std::path::PathBuf;
-#[derive(Debug, clap::Parser)]
-pub struct Cli {
-    #[clap(subcommand)]
-    pub cmd: SubCommand,
-}
-
-#[derive(Debug, clap::Subcommand)]
-pub enum SubCommand {
-    #[clap(name = "detect")]
-    Detect(Detect),
-    #[clap(name = "list")]
-    List(List),
-    #[clap(name = "completions")]
-    Completions { shell: clap_complete::Shell },
-}
-
-#[derive(Debug, clap::ValueEnum, Clone, Copy)]
-pub enum Models {
-    RetinaFace,
-    Yolo,
-}
-
-#[derive(Debug, clap::ValueEnum, Clone, Copy)]
-pub enum Executor {
-    Mnn,
-    Onnx,
-}
-
-#[derive(Debug, clap::ValueEnum, Clone, Copy)]
-pub enum OnnxEp {
-    Cpu,
-}
-
-#[derive(Debug, clap::ValueEnum, Clone, Copy)]
-pub enum MnnEp {
-    Cpu,
-    Metal,
-    OpenCL,
-    CoreML,
-}
-
-#[derive(Debug, clap::Args)]
-pub struct Detect {
-    #[clap(short, long)]
-    pub model: Option<PathBuf>,
-    #[clap(short = 'M', long, default_value = "retina-face")]
-    pub model_type: Models,
-    #[clap(short, long)]
-    pub output: Option<PathBuf>,
-    #[clap(short, long, default_value_t = 0.8)]
-    pub threshold: f32,
-    #[clap(short, long, default_value_t = 0.3)]
-    pub nms_threshold: f32,
-    pub image: PathBuf,
-}
-
-#[derive(Debug, clap::Args)]
-pub struct List {}
-
-impl Cli {
-    pub fn completions(shell: clap_complete::Shell) {
-        let mut command = <Cli as clap::CommandFactory>::command();
-        clap_complete::generate(
-            shell,
-            &mut command,
-            env!("CARGO_BIN_NAME"),
-            &mut std::io::stdout(),
-        );
-    }
-}

src/database.rs

@@ -0,0 +1,734 @@
+use crate::errors::{Error, Result};
+use crate::facedet::{FaceDetectionOutput, FaceLandmarks};
+use bounding_box::Aabb2;
+use error_stack::ResultExt;
+use ndarray_math::CosineSimilarity;
+use rusqlite::{Connection, OptionalExtension, params};
+use std::path::Path;
+
+/// Database connection and operations for face detection results
+pub struct FaceDatabase {
+    conn: Connection,
+}
+
+/// Represents a stored image record
+#[derive(Debug, Clone)]
+pub struct ImageRecord {
+    pub id: i64,
+    pub file_path: String,
+    pub width: u32,
+    pub height: u32,
+    pub created_at: String,
+}
+
+/// Represents a stored face detection record
+#[derive(Debug, Clone)]
+pub struct FaceRecord {
+    pub id: i64,
+    pub image_id: i64,
+    pub bbox_x1: f32,
+    pub bbox_y1: f32,
+    pub bbox_x2: f32,
+    pub bbox_y2: f32,
+    pub confidence: f32,
+    pub created_at: String,
+}
+
+/// Represents stored face landmarks
+#[derive(Debug, Clone)]
+pub struct LandmarkRecord {
+    pub id: i64,
+    pub face_id: i64,
+    pub left_eye_x: f32,
+    pub left_eye_y: f32,
+    pub right_eye_x: f32,
+    pub right_eye_y: f32,
+    pub nose_x: f32,
+    pub nose_y: f32,
+    pub left_mouth_x: f32,
+    pub left_mouth_y: f32,
+    pub right_mouth_x: f32,
+    pub right_mouth_y: f32,
+}
+
+/// Represents a stored face embedding
+#[derive(Debug, Clone)]
+pub struct EmbeddingRecord {
+    pub id: i64,
+    pub face_id: i64,
+    pub embedding: ndarray::Array1<f32>,
+    pub model_name: String,
+    pub created_at: String,
+}
+
+impl FaceDatabase {
+    /// Create a new database connection and initialize tables
+    pub fn new<P: AsRef<Path>>(db_path: P) -> Result<Self> {
+        let conn = Connection::open(db_path).change_context(Error)?;
+        // Temporarily disable extension loading for clustering demo
+        // unsafe {
+        //     let _guard = rusqlite::LoadExtensionGuard::new(&conn).change_context(Error)?;
+        //     conn.load_extension(
+        //         "/Users/fs0c131y/.cache/cargo/target/release/libsqlite3_safetensor_cosine.dylib",
+        //         None::<&str>,
+        //     )
+        //     .change_context(Error)?;
+        // }
+        let db = Self { conn };
+        db.create_tables()?;
+        Ok(db)
+    }
+
+    /// Create an in-memory database for testing
+    pub fn in_memory() -> Result<Self> {
+        let conn = Connection::open_in_memory().change_context(Error)?;
+        let db = Self { conn };
+        db.create_tables()?;
+        Ok(db)
+    }
+
+    /// Create all necessary database tables
+    fn create_tables(&self) -> Result<()> {
+        // Images table
+        self.conn
+            .execute(
+                r#"
+                CREATE TABLE IF NOT EXISTS images (
+                    id INTEGER PRIMARY KEY AUTOINCREMENT,
+                    file_path TEXT NOT NULL UNIQUE,
+                    width INTEGER NOT NULL,
+                    height INTEGER NOT NULL,
+                    created_at DATETIME DEFAULT CURRENT_TIMESTAMP
+                )
+                "#,
+                [],
+            )
+            .change_context(Error)?;
+
+        // Faces table
+        self.conn
+            .execute(
+                r#"
+                CREATE TABLE IF NOT EXISTS faces (
+                    id INTEGER PRIMARY KEY AUTOINCREMENT,
+                    image_id INTEGER NOT NULL,
+                    bbox_x1 REAL NOT NULL,
+                    bbox_y1 REAL NOT NULL,
+                    bbox_x2 REAL NOT NULL,
+                    bbox_y2 REAL NOT NULL,
+                    confidence REAL NOT NULL,
+                    created_at DATETIME DEFAULT CURRENT_TIMESTAMP,
+                    FOREIGN KEY (image_id) REFERENCES images (id) ON DELETE CASCADE
+                )
+                "#,
+                [],
+            )
+            .change_context(Error)?;
+
+        // Landmarks table
+        self.conn
+            .execute(
+                r#"
+                CREATE TABLE IF NOT EXISTS landmarks (
+                    id INTEGER PRIMARY KEY AUTOINCREMENT,
+                    face_id INTEGER NOT NULL,
+                    left_eye_x REAL NOT NULL,
+                    left_eye_y REAL NOT NULL,
+                    right_eye_x REAL NOT NULL,
+                    right_eye_y REAL NOT NULL,
+                    nose_x REAL NOT NULL,
+                    nose_y REAL NOT NULL,
+                    left_mouth_x REAL NOT NULL,
+                    left_mouth_y REAL NOT NULL,
+                    right_mouth_x REAL NOT NULL,
+                    right_mouth_y REAL NOT NULL,
+                    FOREIGN KEY (face_id) REFERENCES faces (id) ON DELETE CASCADE
+                )
+                "#,
+                [],
+            )
+            .change_context(Error)?;
+
+        // Embeddings table
+        self.conn
+            .execute(
+                r#"
+                CREATE TABLE IF NOT EXISTS embeddings (
+                    id INTEGER PRIMARY KEY AUTOINCREMENT,
+                    face_id INTEGER NOT NULL,
+                    embedding BLOB NOT NULL,
+                    model_name TEXT NOT NULL,
+                    created_at DATETIME DEFAULT CURRENT_TIMESTAMP,
+                    FOREIGN KEY (face_id) REFERENCES faces (id) ON DELETE CASCADE
+                )
+                "#,
+                [],
+            )
+            .change_context(Error)?;
+
+        // Create indexes for better performance
+        self.conn
+            .execute(
+                "CREATE INDEX IF NOT EXISTS idx_faces_image_id ON faces (image_id)",
+                [],
+            )
+            .change_context(Error)?;
+
+        self.conn
+            .execute(
+                "CREATE INDEX IF NOT EXISTS idx_landmarks_face_id ON landmarks (face_id)",
+                [],
+            )
+            .change_context(Error)?;
+
+        self.conn
+            .execute(
+                "CREATE INDEX IF NOT EXISTS idx_embeddings_face_id ON embeddings (face_id)",
+                [],
+            )
+            .change_context(Error)?;
+
+        Ok(())
+    }
+
+    /// Store image metadata and return the image ID
+    pub fn store_image(&self, file_path: &str, width: u32, height: u32) -> Result<i64> {
+        let mut stmt = self
+            .conn
+            .prepare("INSERT OR REPLACE INTO images (file_path, width, height) VALUES (?1, ?2, ?3)")
+            .change_context(Error)?;
+
+        Ok(stmt
+            .insert(params![file_path, width, height])
+            .change_context(Error)?)
+    }
+
+    /// Store face detection results
+    pub fn store_face_detections(
+        &self,
+        image_id: i64,
+        detection_output: &FaceDetectionOutput,
+    ) -> Result<Vec<i64>> {
+        let mut face_ids = Vec::new();
+
+        for (i, bbox) in detection_output.bbox.iter().enumerate() {
+            let confidence = detection_output.confidence.get(i).copied().unwrap_or(0.0);
+
+            let face_id = self.store_face(image_id, bbox, confidence)?;
+            face_ids.push(face_id);
+
+            // Store landmarks if available
+            if let Some(landmarks) = detection_output.landmark.get(i) {
+                self.store_landmarks(face_id, landmarks)?;
+            }
+        }
+
+        Ok(face_ids)
+    }
+
+    /// Store a single face detection
+    pub fn store_face(&self, image_id: i64, bbox: &Aabb2<usize>, confidence: f32) -> Result<i64> {
+        let mut stmt = self
+            .conn
+            .prepare(
+                r#"
+                INSERT INTO faces (image_id, bbox_x1, bbox_y1, bbox_x2, bbox_y2, confidence)
+                VALUES (?1, ?2, ?3, ?4, ?5, ?6)
+                "#,
+            )
+            .change_context(Error)?;
+
+        Ok(stmt
+            .insert(params![
+                image_id,
+                bbox.x1() as f32,
+                bbox.y1() as f32,
+                bbox.x2() as f32,
+                bbox.y2() as f32,
+                confidence
+            ])
+            .change_context(Error)?)
+    }
+
+    /// Store face landmarks
+    pub fn store_landmarks(&self, face_id: i64, landmarks: &FaceLandmarks) -> Result<i64> {
+        let mut stmt = self
+            .conn
+            .prepare(
+                r#"
+                INSERT INTO landmarks
+                (face_id, left_eye_x, left_eye_y, right_eye_x, right_eye_y,
+                 nose_x, nose_y, left_mouth_x, left_mouth_y, right_mouth_x, right_mouth_y)
+                VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7, ?8, ?9, ?10, ?11)
+                "#,
+            )
+            .change_context(Error)?;
+
+        Ok(stmt
+            .insert(params![
+                face_id,
+                landmarks.left_eye.x,
+                landmarks.left_eye.y,
+                landmarks.right_eye.x,
+                landmarks.right_eye.y,
+                landmarks.nose.x,
+                landmarks.nose.y,
+                landmarks.left_mouth.x,
+                landmarks.left_mouth.y,
+                landmarks.right_mouth.x,
+                landmarks.right_mouth.y,
+            ])
+            .change_context(Error)?)
+    }
+
+    /// Store face embeddings
+    pub fn store_embeddings(
+        &self,
+        face_ids: &[i64],
+        embeddings: &[ndarray::Array2<f32>],
+        model_name: &str,
+    ) -> Result<Vec<i64>> {
+        let mut embedding_ids = Vec::new();
+
+        for (face_idx, embedding_batch) in embeddings.iter().enumerate() {
+            for (batch_idx, embedding_row) in embedding_batch.rows().into_iter().enumerate() {
+                let global_idx = face_idx * embedding_batch.nrows() + batch_idx;
+
+                if global_idx >= face_ids.len() {
+                    break;
+                }
+
+                let face_id = face_ids[global_idx];
+                let embedding_id =
+                    self.store_single_embedding(face_id, embedding_row, model_name)?;
+                embedding_ids.push(embedding_id);
+            }
+        }
+
+        Ok(embedding_ids)
+    }
+
+    /// Store a single embedding
+    pub fn store_single_embedding(
+        &self,
+        face_id: i64,
+        embedding: ndarray::ArrayView1<f32>,
+        model_name: &str,
+    ) -> Result<i64> {
+        let safe_arrays =
+            ndarray_safetensors::SafeArrays::from_ndarrays([("embedding", embedding)])
+                .change_context(Error)?;
+
+        let embedding_bytes = safe_arrays.serialize().change_context(Error)?;
+
+        let mut stmt = self
+            .conn
+            .prepare("INSERT INTO embeddings (face_id, embedding, model_name) VALUES (?1, ?2, ?3)")
+            .change_context(Error)?;
+
+        stmt.execute(params![face_id, embedding_bytes, model_name])
+            .change_context(Error)?;
+
+        Ok(self.conn.last_insert_rowid())
+    }
+
+    /// Get image by ID
+    pub fn get_image(&self, image_id: i64) -> Result<Option<ImageRecord>> {
+        let mut stmt = self
+            .conn
+            .prepare("SELECT id, file_path, width, height, created_at FROM images WHERE id = ?1")
+            .change_context(Error)?;
+
+        let result = stmt
+            .query_row(params![image_id], |row| {
+                Ok(ImageRecord {
+                    id: row.get(0)?,
+                    file_path: row.get(1)?,
+                    width: row.get(2)?,
+                    height: row.get(3)?,
+                    created_at: row.get(4)?,
+                })
+            })
+            .optional()
+            .change_context(Error)?;
+
+        Ok(result)
+    }
+
+    /// Get all faces for an image
+    pub fn get_faces_for_image(&self, image_id: i64) -> Result<Vec<FaceRecord>> {
+        let mut stmt = self
+            .conn
+            .prepare(
+                r#"
+                SELECT id, image_id, bbox_x1, bbox_y1, bbox_x2, bbox_y2, confidence, created_at
+                FROM faces WHERE image_id = ?1
+                "#,
+            )
+            .change_context(Error)?;
+
+        let face_iter = stmt
+            .query_map(params![image_id], |row| {
+                Ok(FaceRecord {
+                    id: row.get(0)?,
+                    image_id: row.get(1)?,
+                    bbox_x1: row.get(2)?,
+                    bbox_y1: row.get(3)?,
+                    bbox_x2: row.get(4)?,
+                    bbox_y2: row.get(5)?,
+                    confidence: row.get(6)?,
+                    created_at: row.get(7)?,
+                })
+            })
+            .change_context(Error)?;
+
+        let mut faces = Vec::new();
+        for face in face_iter {
+            faces.push(face.change_context(Error)?);
+        }
+
+        Ok(faces)
+    }
+
+    /// Get landmarks for a face
+    pub fn get_landmarks(&self, face_id: i64) -> Result<Option<LandmarkRecord>> {
+        let mut stmt = self
+            .conn
+            .prepare(
+                r#"
+                SELECT id, face_id, left_eye_x, left_eye_y, right_eye_x, right_eye_y,
+                       nose_x, nose_y, left_mouth_x, left_mouth_y, right_mouth_x, right_mouth_y
+                FROM landmarks WHERE face_id = ?1
+                "#,
+            )
+            .change_context(Error)?;
+
+        let result = stmt
+            .query_row(params![face_id], |row| {
+                Ok(LandmarkRecord {
+                    id: row.get(0)?,
+                    face_id: row.get(1)?,
+                    left_eye_x: row.get(2)?,
+                    left_eye_y: row.get(3)?,
+                    right_eye_x: row.get(4)?,
+                    right_eye_y: row.get(5)?,
+                    nose_x: row.get(6)?,
+                    nose_y: row.get(7)?,
+                    left_mouth_x: row.get(8)?,
+                    left_mouth_y: row.get(9)?,
+                    right_mouth_x: row.get(10)?,
+                    right_mouth_y: row.get(11)?,
+                })
+            })
+            .optional()
+            .change_context(Error)?;
+
+        Ok(result)
+    }
+
+    /// Get embeddings for a face
+    pub fn get_embeddings(&self, face_id: i64) -> Result<Vec<EmbeddingRecord>> {
+        let mut stmt = self
+            .conn
+            .prepare(
+                "SELECT id, face_id, embedding, model_name, created_at FROM embeddings WHERE face_id = ?1",
+            )
+            .change_context(Error)?;
+
+        let embedding_iter = stmt
+            .query_map(params![face_id], |row| {
+                let embedding_bytes: Vec<u8> = row.get(2)?;
+                let embedding: ndarray::Array1<f32> = {
+                    let sf = ndarray_safetensors::SafeArraysView::from_bytes(&embedding_bytes)
+                        .change_context(Error)
+                        // .change_context(Error)?
+                        .unwrap();
+                    sf.tensor::<f32, ndarray::Ix1>("embedding")
+                        // .change_context(Error)?
+                        .unwrap()
+                        .to_owned()
+                };
+
+                Ok(EmbeddingRecord {
+                    id: row.get(0)?,
+                    face_id: row.get(1)?,
+                    embedding,
+                    model_name: row.get(3)?,
+                    created_at: row.get(4)?,
+                })
+            })
+            .change_context(Error)?;
+
+        let mut embeddings = Vec::new();
+        for embedding in embedding_iter {
+            embeddings.push(embedding.change_context(Error)?);
+        }
+
+        Ok(embeddings)
+    }
+
+    pub fn get_image_for_face(&self, face_id: i64) -> Result<Option<ImageRecord>> {
+        let mut stmt = self
+            .conn
+            .prepare(
+                r#"
+                SELECT images.id, images.file_path, images.width, images.height, images.created_at
+                FROM images
+                JOIN faces ON faces.image_id = images.id
+                WHERE faces.id = ?1
+                "#,
+            )
+            .change_context(Error)?;
+
+        let result = stmt
+            .query_row(params![face_id], |row| {
+                Ok(ImageRecord {
+                    id: row.get(0)?,
+                    file_path: row.get(1)?,
+                    width: row.get(2)?,
+                    height: row.get(3)?,
+                    created_at: row.get(4)?,
+                })
+            })
+            .optional()
+            .change_context(Error)?;
+
+        Ok(result)
+    }
+
+    /// Get database statistics
+    pub fn get_stats(&self) -> Result<(usize, usize, usize, usize)> {
+        let images: usize = self
+            .conn
+            .query_row("SELECT COUNT(*) FROM images", [], |row| row.get(0))
+            .change_context(Error)?;
+
+        let faces: usize = self
+            .conn
+            .query_row("SELECT COUNT(*) FROM faces", [], |row| row.get(0))
+            .change_context(Error)?;
+
+        let landmarks: usize = self
+            .conn
+            .query_row("SELECT COUNT(*) FROM landmarks", [], |row| row.get(0))
+            .change_context(Error)?;
+
+        let embeddings: usize = self
+            .conn
+            .query_row("SELECT COUNT(*) FROM embeddings", [], |row| row.get(0))
+            .change_context(Error)?;
+
+        Ok((images, faces, landmarks, embeddings))
+    }
+
+    /// Find similar faces based on cosine similarity of embeddings
+    /// Return ids and similarity scores of similar faces
+    pub fn find_similar_faces(
+        &self,
+        embedding: &ndarray::Array1<f32>,
+        threshold: f32,
+        limit: usize,
+    ) -> Result<Vec<(i64, f32)>> {
+        // Serialize the query embedding to bytes
+        let embedding_bytes =
+            ndarray_safetensors::SafeArrays::from_ndarrays([("embedding", embedding.view())])
+                .change_context(Error)?
+                .serialize()
+                .change_context(Error)?;
+
+        let mut stmt = self
+            .conn
+            .prepare(
+                r#" SELECT face_id, cosine_similarity(?1, embedding) as similarity
+                    FROM embeddings
+                    WHERE cosine_similarity(?1, embedding) >= ?2
+                    ORDER BY similarity DESC
+                    LIMIT ?3"#,
+            )
+            .change_context(Error)?;
+
+        let result = stmt
+            .query_map(params![embedding_bytes, threshold, limit], |row| {
+                Ok((row.get::<_, i64>(0)?, row.get::<_, f32>(1)?))
+            })
+            .change_context(Error)?
+            .map(|r| r.change_context(Error))
+            .collect::<Result<Vec<_>>>()?;
+
+        // let mut results = Vec::new();
+        // for result in result_iter {
+        //     results.push(result.change_context(Error)?);
+        // }
+
+        Ok(result)
+    }
+
+    pub fn query_similarity(&self, embedding: &ndarray::Array1<f32>) {
+        let embedding_bytes =
+            ndarray_safetensors::SafeArrays::from_ndarrays([("embedding", embedding.view())])
+                .change_context(Error)
+                .unwrap()
+                .serialize()
+                .change_context(Error)
+                .unwrap();
+
+        let mut stmt = self
+            .conn
+            .prepare(
+                r#"
+                SELECT face_id,
+                cosine_similarity(?1, embedding)
+                FROM embeddings
+                "#,
+            )
+            .change_context(Error)
+            .unwrap();
+
+        let result_iter = stmt
+            .query_map(params![embedding_bytes], |row| {
+                Ok((row.get::<_, i64>(0)?, row.get::<_, f32>(1)?))
+            })
+            .change_context(Error)
+            .unwrap();
+
+        for result in result_iter {
+            println!("{:?}", result);
+        }
+    }
+
+    /// Get all embeddings for a specific model
+    pub fn get_all_embeddings(&self, model_name: Option<&str>) -> Result<Vec<EmbeddingRecord>> {
+        let mut embeddings = Vec::new();
+
+        if let Some(model) = model_name {
+            let mut stmt = self.conn.prepare(
+                "SELECT id, face_id, embedding, model_name, created_at FROM embeddings WHERE model_name = ?1"
+            ).change_context(Error)?;
+
+            let embedding_iter = stmt
+                .query_map(params![model], |row| {
+                    let embedding_bytes: Vec<u8> = row.get(2)?;
+                    let embedding: ndarray::Array1<f32> = {
+                        let sf = ndarray_safetensors::SafeArraysView::from_bytes(&embedding_bytes)
+                            .change_context(Error)
+                            .unwrap();
+                        sf.tensor::<f32, ndarray::Ix1>("embedding")
+                            .unwrap()
+                            .to_owned()
+                    };
+
+                    Ok(EmbeddingRecord {
+                        id: row.get(0)?,
+                        face_id: row.get(1)?,
+                        embedding,
+                        model_name: row.get(3)?,
+                        created_at: row.get(4)?,
+                    })
+                })
+                .change_context(Error)?;
+
+            for embedding in embedding_iter {
+                embeddings.push(embedding.change_context(Error)?);
+            }
+        } else {
+            let mut stmt = self
+                .conn
+                .prepare("SELECT id, face_id, embedding, model_name, created_at FROM embeddings")
+                .change_context(Error)?;
+
+            let embedding_iter = stmt
+                .query_map([], |row| {
+                    let embedding_bytes: Vec<u8> = row.get(2)?;
+                    let embedding: ndarray::Array1<f32> = {
+                        let sf = ndarray_safetensors::SafeArraysView::from_bytes(&embedding_bytes)
+                            .change_context(Error)
+                            .unwrap();
+                        sf.tensor::<f32, ndarray::Ix1>("embedding")
+                            .unwrap()
+                            .to_owned()
+                    };
+
+                    Ok(EmbeddingRecord {
+                        id: row.get(0)?,
+                        face_id: row.get(1)?,
+                        embedding,
+                        model_name: row.get(3)?,
+                        created_at: row.get(4)?,
+                    })
+                })
+                .change_context(Error)?;
+
+            for embedding in embedding_iter {
+                embeddings.push(embedding.change_context(Error)?);
+            }
+        }
+
+        Ok(embeddings)
+    }
+}
+
+fn add_sqlite_cosine_similarity(db: &Connection) -> Result<()> {
+    use rusqlite::functions::*;
+    db.create_scalar_function(
+        "cosine_similarity",
+        2,
+        FunctionFlags::SQLITE_UTF8 | FunctionFlags::SQLITE_DETERMINISTIC,
+        move |ctx| {
+            if ctx.len() != 2 {
+                return Err(rusqlite::Error::UserFunctionError(
+                    "cosine_similarity requires exactly 2 arguments".into(),
+                ));
+            }
+            let array_1 = ctx.get_raw(0).as_blob()?;
+            let array_2 = ctx.get_raw(1).as_blob()?;
+
+            let array_1_st = ndarray_safetensors::SafeArraysView::from_bytes(array_1)
+                .map_err(|e| rusqlite::Error::UserFunctionError(e.into()))?;
+            let array_2_st = ndarray_safetensors::SafeArraysView::from_bytes(array_2)
+                .map_err(|e| rusqlite::Error::UserFunctionError(e.into()))?;
+
+            let array_view_1 = array_1_st
+                .tensor_by_index::<f32, ndarray::Ix1>(0)
+                .map_err(|e| rusqlite::Error::UserFunctionError(e.into()))?;
+            let array_view_2 = array_2_st
+                .tensor_by_index::<f32, ndarray::Ix1>(0)
+                .map_err(|e| rusqlite::Error::UserFunctionError(e.into()))?;
+
+            let similarity = array_view_1
+                .cosine_similarity(array_view_2)
+                .map_err(|e| rusqlite::Error::UserFunctionError(e.into()))?;
+
+            Ok(similarity)
+        },
+    )
+    .change_context(Error)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_database_creation() -> Result<()> {
+        let db = FaceDatabase::in_memory()?;
+        let (images, faces, landmarks, embeddings) = db.get_stats()?;
+        assert_eq!(images, 0);
+        assert_eq!(faces, 0);
+        assert_eq!(landmarks, 0);
+        assert_eq!(embeddings, 0);
+        Ok(())
+    }
+
+    #[test]
+    fn test_store_and_retrieve_image() -> Result<()> {
+        let db = FaceDatabase::in_memory()?;
+        let image_id = db.store_image("/path/to/image.jpg", 800, 600)?;
+
+        let image = db.get_image(image_id)?.unwrap();
+        assert_eq!(image.file_path, "/path/to/image.jpg");
+        assert_eq!(image.width, 800);
+        assert_eq!(image.height, 600);
+
+        Ok(())
+    }
+}

src/facedet.rs

@@ -1,2 +1,8 @@
 pub mod retinaface;
 pub mod yolo;
+
+// Re-export common types and traits
+pub use retinaface::{
+    FaceDetectionConfig, FaceDetectionModelOutput, FaceDetectionOutput,
+    FaceDetectionProcessedOutput, FaceDetector, FaceLandmarks,
+};

src/facedet/retinaface.rs

@@ -1,67 +1,88 @@
+pub mod mnn;
+pub mod ort;
+
 use crate::errors::*;
 use bounding_box::{Aabb2, nms::nms};
 use error_stack::ResultExt;
-use mnn_bridge::ndarray::*;
 use nalgebra::{Point2, Vector2};
-use ndarray_resize::NdFir;
-use std::path::Path;
 
+/// Configuration for face detection postprocessing
+#[derive(Debug, Clone, PartialEq)]
 pub struct FaceDetectionConfig {
-    anchor_sizes: Vec<Vector2<usize>>,
-    steps: Vec<usize>,
-    variance: Vec<f32>,
-    threshold: f32,
-    nms_threshold: f32,
+    /// Minimum confidence to keep a detection
+    pub threshold: f32,
+    /// NMS threshold for suppressing overlapping boxes
+    pub nms_threshold: f32,
+    /// Variances for bounding box decoding
+    pub variances: [f32; 2],
+    /// The step size (stride) for each feature map
+    pub steps: Vec<usize>,
+    /// The minimum anchor sizes for each feature map
+    pub min_sizes: Vec<Vec<usize>>,
+    /// Whether to clip bounding boxes to the image dimensions
+    pub clamp: bool,
+    /// Input image width (used for anchor generation)
+    pub input_width: usize,
+    /// Input image height (used for anchor generation)
+    pub input_height: usize,
 }
 
 impl FaceDetectionConfig {
-    pub fn with_min_sizes(mut self, min_sizes: Vec<Vector2<usize>>) -> Self {
-        self.anchor_sizes = min_sizes;
-        self
-    }
-    pub fn with_steps(mut self, steps: Vec<usize>) -> Self {
-        self.steps = steps;
-        self
-    }
-    pub fn with_variance(mut self, variance: Vec<f32>) -> Self {
-        self.variance = variance;
-        self
-    }
     pub fn with_threshold(mut self, threshold: f32) -> Self {
         self.threshold = threshold;
         self
     }
+
     pub fn with_nms_threshold(mut self, nms_threshold: f32) -> Self {
         self.nms_threshold = nms_threshold;
         self
     }
+
+    pub fn with_variances(mut self, variances: [f32; 2]) -> Self {
+        self.variances = variances;
+        self
+    }
+
+    pub fn with_steps(mut self, steps: Vec<usize>) -> Self {
+        self.steps = steps;
+        self
+    }
+
+    pub fn with_min_sizes(mut self, min_sizes: Vec<Vec<usize>>) -> Self {
+        self.min_sizes = min_sizes;
+        self
+    }
+
+    pub fn with_clip(mut self, clip: bool) -> Self {
+        self.clamp = clip;
+        self
+    }
+
+    pub fn with_input_width(mut self, input_width: usize) -> Self {
+        self.input_width = input_width;
+        self
+    }
+
+    pub fn with_input_height(mut self, input_height: usize) -> Self {
+        self.input_height = input_height;
+        self
+    }
 }
 
 impl Default for FaceDetectionConfig {
     fn default() -> Self {
-        FaceDetectionConfig {
-            anchor_sizes: vec![
-                Vector2::new(16, 32),
-                Vector2::new(64, 128),
-                Vector2::new(256, 512),
-            ],
-            steps: vec![8, 16, 32],
-            variance: vec![0.1, 0.2],
-            threshold: 0.8,
+        Self {
+            threshold: 0.5,
             nms_threshold: 0.4,
+            variances: [0.1, 0.2],
+            steps: vec![8, 16, 32],
+            min_sizes: vec![vec![16, 32], vec![64, 128], vec![256, 512]],
+            clamp: true,
+            input_width: 1024,
+            input_height: 1024,
         }
     }
 }
-pub struct FaceDetection {
-    handle: mnn_sync::SessionHandle,
-}
-
-#[derive(Debug, Clone, PartialEq)]
-pub struct FaceDetectionModelOutput {
-    pub bbox: ndarray::Array3<f32>,
-    pub confidence: ndarray::Array3<f32>,
-    pub landmark: ndarray::Array3<f32>,
-}
 
 /// Represents the 5 facial landmarks detected by RetinaFace
 #[derive(Debug, Copy, Clone, PartialEq)]
@@ -73,6 +94,13 @@ pub struct FaceLandmarks {
     pub right_mouth: Point2<f32>,
 }
 
+#[derive(Debug, Clone, PartialEq)]
+pub struct FaceDetectionModelOutput {
+    pub bbox: ndarray::Array3<f32>,
+    pub confidence: ndarray::Array3<f32>,
+    pub landmark: ndarray::Array3<f32>,
+}
+
 #[derive(Debug, Clone, PartialEq)]
 pub struct FaceDetectionProcessedOutput {
     pub bbox: Vec<Aabb2<f32>>,
@@ -87,85 +115,133 @@ pub struct FaceDetectionOutput {
     pub landmark: Vec<FaceLandmarks>,
 }
 
-impl FaceDetectionModelOutput {
-    pub fn postprocess(self, config: &FaceDetectionConfig) -> Result<FaceDetectionProcessedOutput> {
+/// Raw model outputs that can be converted to FaceDetectionModelOutput
+pub trait IntoModelOutput {
+    fn into_model_output(self) -> Result<FaceDetectionModelOutput>;
+}
+
+/// Generate anchors for RetinaFace model
+pub fn generate_anchors(config: &FaceDetectionConfig) -> ndarray::Array2<f32> {
     let mut anchors = Vec::new();
-        for (k, &step) in config.steps.iter().enumerate() {
-            let feature_size = 1024 / step;
-            let min_sizes = config.anchor_sizes[k];
-            let sizes = [min_sizes.x, min_sizes.y];
-            for i in 0..feature_size {
-                for j in 0..feature_size {
-                    for &size in &sizes {
-                        let cx = (j as f32 + 0.5) * step as f32 / 1024.0;
-                        let cy = (i as f32 + 0.5) * step as f32 / 1024.0;
-                        let s_k = size as f32 / 1024.0;
-                        anchors.push((cx, cy, s_k, s_k));
-                    }
-                }
-            }
-        }
-        let mut boxes = Vec::new();
-        let mut scores = Vec::new();
-        let mut landmarks = Vec::new();
-        let var0 = config.variance[0];
-        let var1 = config.variance[1];
-        let bbox_data = self.bbox;
-        let conf_data = self.confidence;
-        let landmark_data = self.landmark;
-        let num_priors = bbox_data.shape()[1];
-        for idx in 0..num_priors {
-            let dx = bbox_data[[0, idx, 0]];
-            let dy = bbox_data[[0, idx, 1]];
-            let dw = bbox_data[[0, idx, 2]];
-            let dh = bbox_data[[0, idx, 3]];
-            let (anchor_cx, anchor_cy, anchor_w, anchor_h) = anchors[idx];
-            let pred_cx = anchor_cx + dx * var0 * anchor_w;
-            let pred_cy = anchor_cy + dy * var0 * anchor_h;
-            let pred_w = anchor_w * (dw * var1).exp();
-            let pred_h = anchor_h * (dh * var1).exp();
-            let x_min = pred_cx - pred_w / 2.0;
-            let y_min = pred_cy - pred_h / 2.0;
-            let x_max = pred_cx + pred_w / 2.0;
-            let y_max = pred_cy + pred_h / 2.0;
-            let score = conf_data[[0, idx, 1]];
-            if score > config.threshold {
-                boxes.push(Aabb2::from_x1y1x2y2(x_min, y_min, x_max, y_max));
-                scores.push(score);
-
-                let left_eye_x = landmark_data[[0, idx, 0]] * anchor_w * var0 + anchor_cx;
-                let left_eye_y = landmark_data[[0, idx, 1]] * anchor_h * var0 + anchor_cy;
-
-                let right_eye_x = landmark_data[[0, idx, 2]] * anchor_w * var0 + anchor_cx;
-                let right_eye_y = landmark_data[[0, idx, 3]] * anchor_h * var0 + anchor_cy;
-
-                let nose_x = landmark_data[[0, idx, 4]] * anchor_w * var0 + anchor_cx;
-                let nose_y = landmark_data[[0, idx, 5]] * anchor_h * var0 + anchor_cy;
-
-                let left_mouth_x = landmark_data[[0, idx, 6]] * anchor_w * var0 + anchor_cx;
-                let left_mouth_y = landmark_data[[0, idx, 7]] * anchor_h * var0 + anchor_cy;
-
-                let right_mouth_x = landmark_data[[0, idx, 8]] * anchor_w * var0 + anchor_cx;
-                let right_mouth_y = landmark_data[[0, idx, 9]] * anchor_h * var0 + anchor_cy;
-
-                landmarks.push(FaceLandmarks {
-                    left_eye: Point2::new(left_eye_x, left_eye_y),
-                    right_eye: Point2::new(right_eye_x, right_eye_y),
-                    nose: Point2::new(nose_x, nose_y),
-                    left_mouth: Point2::new(left_mouth_x, left_mouth_y),
-                    right_mouth: Point2::new(right_mouth_x, right_mouth_y),
-                });
-            }
-        }
-        Ok(FaceDetectionProcessedOutput {
-            bbox: boxes,
-            confidence: scores,
-            landmarks,
+    let feature_maps: Vec<(usize, usize)> = config
+        .steps
+        .iter()
+        .map(|&step| {
+            (
+                (config.input_height as f32 / step as f32).ceil() as usize,
+                (config.input_width as f32 / step as f32).ceil() as usize,
+            )
         })
+        .collect();
+
+    for (k, f) in feature_maps.iter().enumerate() {
+        let min_sizes = &config.min_sizes[k];
+        for i in 0..f.0 {
+            for j in 0..f.1 {
+                for &min_size in min_sizes {
+                    let s_kx = min_size as f32 / config.input_width as f32;
+                    let s_ky = min_size as f32 / config.input_height as f32;
+                    let dense_cx =
+                        (j as f32 + 0.5) * config.steps[k] as f32 / config.input_width as f32;
+                    let dense_cy =
+                        (i as f32 + 0.5) * config.steps[k] as f32 / config.input_height as f32;
+                    anchors.push([
+                        dense_cx - s_kx / 2.,
+                        dense_cy - s_ky / 2.,
+                        dense_cx + s_kx / 2.,
+                        dense_cy + s_ky / 2.,
+                    ]);
                 }
+            }
+        }
+    }
+
+    ndarray::Array2::from_shape_vec((anchors.len(), 4), anchors.into_iter().flatten().collect())
+        .unwrap()
 }
 
 impl FaceDetectionModelOutput {
+    pub fn postprocess(self, config: &FaceDetectionConfig) -> Result<FaceDetectionProcessedOutput> {
+        use ndarray::s;
+
+        let priors = generate_anchors(config);
+
+        let scores = self.confidence.slice(s![0, .., 1]);
+        let boxes = self.bbox.slice(s![0, .., ..]);
+        let landmarks_raw = self.landmark.slice(s![0, .., ..]);
+
+        // let mut decoded_boxes = Vec::new();
+        // let mut decoded_landmarks = Vec::new();
+        // let mut confidences = Vec::new();
+
+        let (decoded_boxes, decoded_landmarks, confidences) = (0..priors.shape()[0])
+            .filter(|&i| scores[i] > config.threshold)
+            .map(|i| {
+                let prior = priors.row(i);
+                let loc = boxes.row(i);
+                let landm = landmarks_raw.row(i);
+
+                // Decode bounding box
+                let prior_cx = (prior[0] + prior[2]) / 2.0;
+                let prior_cy = (prior[1] + prior[3]) / 2.0;
+                let prior_w = prior[2] - prior[0];
+                let prior_h = prior[3] - prior[1];
+
+                let var = config.variances;
+                let cx = prior_cx + loc[0] * var[0] * prior_w;
+                let cy = prior_cy + loc[1] * var[0] * prior_h;
+                let w = prior_w * (loc[2] * var[1]).exp();
+                let h = prior_h * (loc[3] * var[1]).exp();
+
+                let xmin = cx - w / 2.0;
+                let ymin = cy - h / 2.0;
+                let xmax = cx + w / 2.0;
+                let ymax = cy + h / 2.0;
+
+                let mut bbox =
+                    Aabb2::from_min_max_vertices(Point2::new(xmin, ymin), Point2::new(xmax, ymax));
+                if config.clamp {
+                    bbox = bbox.component_clamp(0.0, 1.0);
+                }
+
+                // Decode landmarks
+                let points: [Point2<f32>; 5] = (0..5)
+                    .map(|j| {
+                        Point2::new(
+                            prior_cx + landm[j * 2] * var[0] * prior_w,
+                            prior_cy + landm[j * 2 + 1] * var[0] * prior_h,
+                        )
+                    })
+                    .collect::<Vec<_>>()
+                    .try_into()
+                    .unwrap();
+
+                let landmarks = FaceLandmarks {
+                    left_eye: points[0],
+                    right_eye: points[1],
+                    nose: points[2],
+                    left_mouth: points[3],
+                    right_mouth: points[4],
+                };
+
+                (bbox, landmarks, scores[i])
+            })
+            .fold(
+                (Vec::new(), Vec::new(), Vec::new()),
+                |(mut boxes, mut landmarks, mut confs), (bbox, landmark, conf)| {
+                    boxes.push(bbox);
+                    landmarks.push(landmark);
+                    confs.push(conf);
+                    (boxes, landmarks, confs)
+                },
+            );
+        Ok(FaceDetectionProcessedOutput {
+            bbox: decoded_boxes,
+            confidence: confidences,
+            landmarks: decoded_landmarks,
+        })
+    }
+
     pub fn print(&self, limit: usize) {
         tracing::info!("Detected {} faces", self.bbox.shape()[1]);
 
@@ -189,49 +265,16 @@ impl FaceDetectionModelOutput {
     }
 }
 
-impl FaceDetection {
-    pub fn new(path: impl AsRef<Path>) -> Result<Self> {
-        let model = std::fs::read(path)
-            .change_context(Error)
-            .attach_printable("Failed to read model file")?;
-        Self::new_from_bytes(&model)
-    }
-
-    pub fn new_from_bytes(model: &[u8]) -> Result<Self> {
-        tracing::info!("Loading face detection model from bytes");
-        let mut model = mnn::Interpreter::from_bytes(model)
-            .map_err(|e| e.into_inner())
-            .change_context(Error)
-            .attach_printable("Failed to load model from bytes")?;
-        model.set_session_mode(mnn::SessionMode::Release);
-        let bc = mnn::BackendConfig::default().with_memory_mode(mnn::MemoryMode::High);
-        let sc = mnn::ScheduleConfig::new()
-            .with_type(mnn::ForwardType::CPU)
-            .with_backend_config(bc);
-        tracing::info!("Creating session handle for face detection model");
-        let handle = mnn_sync::SessionHandle::new(model, sc)
-            .change_context(Error)
-            .attach_printable("Failed to create session handle")?;
-        Ok(FaceDetection { handle })
-    }
-
-    pub fn detect_faces(
-        &self,
-        image: ndarray::Array3<u8>,
-        config: FaceDetectionConfig,
-    ) -> Result<FaceDetectionOutput> {
-        let (height, width, _channels) = image.dim();
-        let output = self
-            .run_models(image)
-            .change_context(Error)
-            .attach_printable("Failed to detect faces")?;
-        // denormalize the bounding boxes
-        let factor = Vector2::new(width as f32, height as f32);
-        let mut processed = output
-            .postprocess(&config)
-            .attach_printable("Failed to postprocess")?;
-
+/// Apply Non-Maximum Suppression and convert to final output format
+pub fn apply_nms_and_finalize(
+    processed: FaceDetectionProcessedOutput,
+    config: &FaceDetectionConfig,
+    image_size: (usize, usize), // (width, height)
+) -> Result<FaceDetectionOutput> {
     use itertools::Itertools;
+
+    let factor = Vector2::new(image_size.0 as f32, image_size.1 as f32);
+
     let (boxes, scores, landmarks): (Vec<_>, Vec<_>, Vec<_>) = processed
         .bbox
         .iter()
@@ -242,7 +285,8 @@ impl FaceDetection {
         .map(|((b, s), l)| (b, s, l))
         .multiunzip();
 
-        let keep_indices = nms(&boxes, &scores, config.threshold, config.nms_threshold);
+    let keep_indices =
+        nms(&boxes, &scores, config.threshold, config.nms_threshold).change_context(Error)?;
 
     let bboxes = boxes
         .into_iter()
@@ -268,81 +312,29 @@ impl FaceDetection {
         confidence,
         landmark,
     })
-    }
+}
 
-    pub fn run_models(&self, image: ndarray::Array3<u8>) -> Result<FaceDetectionModelOutput> {
-        #[rustfmt::skip]
-        use ::tap::*;
+/// Common trait for face detection backends
+pub trait FaceDetector {
+    /// Run inference on the model and return raw outputs
+    fn run_model(&mut self, image: ndarray::ArrayView3<u8>) -> Result<FaceDetectionModelOutput>;
+
+    /// Detect faces with full pipeline including postprocessing
+    fn detect_faces(
+        &mut self,
+        image: ndarray::ArrayView3<u8>,
+        config: &FaceDetectionConfig,
+    ) -> Result<FaceDetectionOutput> {
+        let (height, width, _channels) = image.dim();
         let output = self
-            .handle
-            .run(move |sr| {
-                let mut resized = image
-                    .fast_resize(1024, 1024, None)
-                    .change_context(mnn::ErrorKind::TensorError)?
-                    .mapv(|f| f as f32)
-                    .tap_mut(|arr| {
-                        arr.axis_iter_mut(ndarray::Axis(2))
-                            .zip([104, 117, 123])
-                            .for_each(|(mut array, pixel)| {
-                                let pixel = pixel as f32;
-                                array.map_inplace(|v| *v -= pixel);
-                            });
-                    })
-                    .permuted_axes((2, 0, 1))
-                    .insert_axis(ndarray::Axis(0))
-                    .as_standard_layout()
-                    .into_owned();
-                let tensor = resized
-                    .as_mnn_tensor_mut()
-                    .attach_printable("Failed to convert ndarray to mnn tensor")
-                    .change_context(mnn::error::ErrorKind::TensorError)?;
-                tracing::trace!("Image Tensor shape: {:?}", tensor.shape());
-                let (intptr, session) = sr.both_mut();
-                tracing::trace!("Copying input tensor to host");
-                unsafe {
-                    let mut input = intptr.input_unresized::<f32>(session, "input")?;
-                    tracing::trace!("Input shape: {:?}", input.shape());
-                    intptr.resize_tensor_by_nchw::<mnn::View<&mut f32>, _>(
-                        input.view_mut(),
-                        1,
-                        3,
-                        1024,
-                        1024,
-                    );
-                }
-                intptr.resize_session(session);
-                let mut input = intptr.input::<f32>(session, "input")?;
-                tracing::trace!("Input shape: {:?}", input.shape());
-                input.copy_from_host_tensor(tensor.view())?;
+            .run_model(image)
+            .change_context(Error)
+            .attach_printable("Failed to detect faces")?;
 
-                tracing::info!("Running face detection session");
-                intptr.run_session(&session)?;
-                let output_tensor = intptr
-                    .output::<f32>(&session, "bbox")?
-                    .create_host_tensor_from_device(true)
-                    .as_ndarray()
-                    .to_owned();
-                tracing::trace!("Output Bbox: \t\t{:?}", output_tensor.shape());
-                let output_confidence = intptr
-                    .output::<f32>(&session, "confidence")?
-                    .create_host_tensor_from_device(true)
-                    .as_ndarray::<ndarray::Ix3>()
-                    .to_owned();
-                tracing::trace!("Output Confidence: \t{:?}", output_confidence.shape());
-                let output_landmark = intptr
-                    .output::<f32>(&session, "landmark")?
-                    .create_host_tensor_from_device(true)
-                    .as_ndarray::<ndarray::Ix3>()
-                    .to_owned();
-                tracing::trace!("Output Landmark: \t{:?}", output_landmark.shape());
-                Ok(FaceDetectionModelOutput {
-                    bbox: output_tensor,
-                    confidence: output_confidence,
-                    landmark: output_landmark,
-                })
-            })
-            .map_err(|e| e.into_inner())
-            .change_context(Error)?;
-        Ok(output)
+        let processed = output
+            .postprocess(&config)
+            .attach_printable("Failed to postprocess")?;
+
+        apply_nms_and_finalize(processed, &config, (width, height))
     }
 }

src/facedet/retinaface/mnn.rs

@@ -0,0 +1,146 @@
+use crate::errors::*;
+use crate::facedet::*;
+use error_stack::ResultExt;
+use mnn_bridge::ndarray::*;
+use ndarray_resize::NdFir;
+use std::path::Path;
+
+#[derive(Debug)]
+pub struct FaceDetection {
+    handle: mnn_sync::SessionHandle,
+}
+
+pub struct FaceDetectionBuilder {
+    schedule_config: Option<mnn::ScheduleConfig>,
+    backend_config: Option<mnn::BackendConfig>,
+    model: mnn::Interpreter,
+}
+
+impl FaceDetectionBuilder {
+    pub fn new(model: impl AsRef<[u8]>) -> Result<Self> {
+        Ok(Self {
+            schedule_config: None,
+            backend_config: None,
+            model: mnn::Interpreter::from_bytes(model.as_ref())
+                .map_err(|e| e.into_inner())
+                .change_context(Error)
+                .attach_printable("Failed to load model from bytes")?,
+        })
+    }
+
+    pub fn with_forward_type(mut self, forward_type: mnn::ForwardType) -> Self {
+        self.schedule_config
+            .get_or_insert_default()
+            .set_type(forward_type);
+        self
+    }
+
+    pub fn with_schedule_config(mut self, config: mnn::ScheduleConfig) -> Self {
+        self.schedule_config = Some(config);
+        self
+    }
+
+    pub fn with_backend_config(mut self, config: mnn::BackendConfig) -> Self {
+        self.backend_config = Some(config);
+        self
+    }
+
+    pub fn build(self) -> Result<FaceDetection> {
+        let model = self.model;
+        let sc = self.schedule_config.unwrap_or_default();
+        let handle = mnn_sync::SessionHandle::new(model, sc)
+            .change_context(Error)
+            .attach_printable("Failed to create session handle")?;
+        Ok(FaceDetection { handle })
+    }
+}
+
+impl FaceDetection {
+    pub fn builder<T: AsRef<[u8]>>(
+        model: T,
+    ) -> std::result::Result<FaceDetectionBuilder, error_stack::Report<crate::errors::Error>> {
+        FaceDetectionBuilder::new(model)
+    }
+}
+
+impl FaceDetector for FaceDetection {
+    fn run_model(&mut self, image: ndarray::ArrayView3<u8>) -> Result<FaceDetectionModelOutput> {
+        #[rustfmt::skip]
+        let mut resized = image
+            .fast_resize(1024, 1024, None)
+            .change_context(Error)?
+            .mapv(|f| f as f32);
+
+        // Apply mean subtraction: [104, 117, 123]
+        resized
+            .axis_iter_mut(ndarray::Axis(2))
+            .zip([104, 117, 123])
+            .for_each(|(mut array, pixel)| {
+                let pixel = pixel as f32;
+                array.map_inplace(|v| *v -= pixel);
+            });
+
+        let mut resized = resized
+            .permuted_axes((2, 0, 1))
+            .insert_axis(ndarray::Axis(0))
+            .as_standard_layout()
+            .into_owned();
+
+        use ::tap::*;
+        let output = self
+            .handle
+            .run(move |sr| {
+                let tensor = resized
+                    .as_mnn_tensor_mut()
+                    .attach_printable("Failed to convert ndarray to mnn tensor")
+                    .change_context(mnn::error::ErrorKind::TensorError)?;
+                tracing::trace!("Image Tensor shape: {:?}", tensor.shape());
+                let (intptr, session) = sr.both_mut();
+                tracing::trace!("Copying input tensor to host");
+                unsafe {
+                    let mut input = intptr.input_unresized::<f32>(session, "input")?;
+                    tracing::trace!("Input shape: {:?}", input.shape());
+                    intptr.resize_tensor_by_nchw::<mnn::View<&mut f32>, _>(
+                        input.view_mut(),
+                        1,
+                        3,
+                        1024,
+                        1024,
+                    );
+                }
+                intptr.resize_session(session);
+                let mut input = intptr.input::<f32>(session, "input")?;
+                tracing::trace!("Input shape: {:?}", input.shape());
+                input.copy_from_host_tensor(tensor.view())?;
+
+                tracing::info!("Running face detection session");
+                intptr.run_session(&session)?;
+                let output_tensor = intptr
+                    .output::<f32>(&session, "bbox")?
+                    .create_host_tensor_from_device(true)
+                    .as_ndarray()
+                    .to_owned();
+                tracing::trace!("Output Bbox: \t\t{:?}", output_tensor.shape());
+                let output_confidence = intptr
+                    .output::<f32>(&session, "confidence")?
+                    .create_host_tensor_from_device(true)
+                    .as_ndarray::<ndarray::Ix3>()
+                    .to_owned();
+                tracing::trace!("Output Confidence: \t{:?}", output_confidence.shape());
+                let output_landmark = intptr
+                    .output::<f32>(&session, "landmark")?
+                    .create_host_tensor_from_device(true)
+                    .as_ndarray::<ndarray::Ix3>()
+                    .to_owned();
+                tracing::trace!("Output Landmark: \t{:?}", output_landmark.shape());
+                Ok(FaceDetectionModelOutput {
+                    bbox: output_tensor,
+                    confidence: output_confidence,
+                    landmark: output_landmark,
+                })
+            })
+            .map_err(|e| e.into_inner())
+            .change_context(Error)?;
+        Ok(output)
+    }
+}

src/facedet/retinaface/ort.rs

@@ -0,0 +1,256 @@
+use crate::errors::*;
+use crate::facedet::*;
+use crate::ort_ep::*;
+use error_stack::ResultExt;
+use ndarray_resize::NdFir;
+use ort::{
+    execution_providers::{CPUExecutionProvider, ExecutionProviderDispatch},
+    session::{Session, builder::GraphOptimizationLevel},
+    value::Tensor,
+};
+use std::path::Path;
+
+#[derive(Debug)]
+pub struct FaceDetection {
+    session: Session,
+}
+
+pub struct FaceDetectionBuilder {
+    model_data: Vec<u8>,
+    execution_providers: Option<Vec<ExecutionProviderDispatch>>,
+    intra_threads: Option<usize>,
+    inter_threads: Option<usize>,
+}
+
+impl FaceDetectionBuilder {
+    pub fn new(model: impl AsRef<[u8]>) -> crate::errors::Result<Self> {
+        Ok(Self {
+            model_data: model.as_ref().to_vec(),
+            execution_providers: None,
+            intra_threads: None,
+            inter_threads: None,
+        })
+    }
+
+    pub fn with_execution_providers(mut self, providers: impl AsRef<[ExecutionProvider]>) -> Self {
+        let execution_providers: Vec<ExecutionProviderDispatch> = providers
+            .as_ref()
+            .iter()
+            .filter_map(|provider| provider.to_dispatch())
+            .collect();
+
+        if !execution_providers.is_empty() {
+            self.execution_providers = Some(execution_providers);
+        } else {
+            tracing::warn!("No valid execution providers found, falling back to CPU");
+            self.execution_providers = Some(vec![CPUExecutionProvider::default().build()]);
+        }
+        self
+    }
+
+    pub fn with_intra_threads(mut self, threads: usize) -> Self {
+        self.intra_threads = Some(threads);
+        self
+    }
+
+    pub fn with_inter_threads(mut self, threads: usize) -> Self {
+        self.inter_threads = Some(threads);
+        self
+    }
+
+    pub fn build(self) -> crate::errors::Result<FaceDetection> {
+        let mut session_builder = Session::builder()
+            .change_context(Error)
+            .attach_printable("Failed to create session builder")?;
+
+        // Set execution providers
+        if let Some(providers) = self.execution_providers {
+            session_builder = session_builder
+                .with_execution_providers(providers)
+                .change_context(Error)
+                .attach_printable("Failed to set execution providers")?;
+        } else {
+            // Default to CPU
+            session_builder = session_builder
+                .with_execution_providers([CPUExecutionProvider::default().build()])
+                .change_context(Error)
+                .attach_printable("Failed to set default CPU execution provider")?;
+        }
+
+        // Set threading options
+        if let Some(threads) = self.intra_threads {
+            session_builder = session_builder
+                .with_intra_threads(threads)
+                .change_context(Error)
+                .attach_printable("Failed to set intra threads")?;
+        }
+
+        if let Some(threads) = self.inter_threads {
+            session_builder = session_builder
+                .with_inter_threads(threads)
+                .change_context(Error)
+                .attach_printable("Failed to set inter threads")?;
+        }
+
+        // Set optimization level
+        session_builder = session_builder
+            .with_optimization_level(GraphOptimizationLevel::Level3)
+            .change_context(Error)
+            .attach_printable("Failed to set optimization level")?;
+
+        // Create session from model bytes
+        let session = session_builder
+            .commit_from_memory(&self.model_data)
+            .change_context(Error)
+            .attach_printable("Failed to create ORT session from model bytes")?;
+
+        tracing::info!("Successfully created ORT RetinaFace session");
+
+        Ok(FaceDetection { session })
+    }
+}
+
+impl FaceDetection {
+    pub fn builder<T: AsRef<[u8]>>(
+        model: T,
+    ) -> std::result::Result<FaceDetectionBuilder, error_stack::Report<crate::errors::Error>> {
+        FaceDetectionBuilder::new(model)
+    }
+
+    pub fn new(path: impl AsRef<Path>) -> crate::errors::Result<Self> {
+        let model = std::fs::read(path)
+            .change_context(Error)
+            .attach_printable("Failed to read model file")?;
+        Self::new_from_bytes(&model)
+    }
+
+    pub fn new_from_bytes(model: &[u8]) -> crate::errors::Result<Self> {
+        tracing::info!("Loading ORT RetinaFace model from bytes");
+        Self::builder(model)?.build()
+    }
+}
+
+impl FaceDetector for FaceDetection {
+    fn run_model(
+        &mut self,
+        image: ndarray::ArrayView3<u8>,
+    ) -> crate::errors::Result<FaceDetectionModelOutput> {
+        // Resize image to 1024x1024
+        let mut resized = image
+            .fast_resize(1024, 1024, None)
+            .change_context(Error)
+            .attach_printable("Failed to resize image")?
+            .mapv(|f| f as f32);
+
+        // Apply mean subtraction: [104, 117, 123] for BGR format
+        resized
+            .axis_iter_mut(ndarray::Axis(2))
+            .zip([104.0, 117.0, 123.0])
+            .for_each(|(mut array, mean)| {
+                array.map_inplace(|v| *v -= mean);
+            });
+
+        // Convert from HWC to NCHW format (add batch dimension and transpose)
+        let input_tensor = resized
+            .permuted_axes((2, 0, 1))
+            .insert_axis(ndarray::Axis(0))
+            .as_standard_layout()
+            .into_owned();
+
+        tracing::trace!("Input tensor shape: {:?}", input_tensor.shape());
+
+        // Create ORT input tensor
+        let input_value = Tensor::from_array(input_tensor)
+            .change_context(Error)
+            .attach_printable("Failed to create input tensor")?;
+
+        // Run inference
+        tracing::debug!("Running ORT RetinaFace inference");
+        let outputs = self
+            .session
+            .run(ort::inputs!["input" => input_value])
+            .change_context(Error)
+            .attach_printable("Failed to run inference")?;
+
+        // Extract outputs by name
+        let bbox_output = outputs
+            .get("bbox")
+            .ok_or(Error)
+            .attach_printable("Missing bbox output from model")?
+            .try_extract_tensor::<f32>()
+            .change_context(Error)
+            .attach_printable("Failed to extract bbox tensor")?;
+
+        let confidence_output = outputs
+            .get("confidence")
+            .ok_or(Error)
+            .attach_printable("Missing confidence output from model")?
+            .try_extract_tensor::<f32>()
+            .change_context(Error)
+            .attach_printable("Failed to extract confidence tensor")?;
+
+        let landmark_output = outputs
+            .get("landmark")
+            .ok_or(Error)
+            .attach_printable("Missing landmark output from model")?
+            .try_extract_tensor::<f32>()
+            .change_context(Error)
+            .attach_printable("Failed to extract landmark tensor")?;
+
+        // Get tensor shapes and data
+        let (bbox_shape, bbox_data) = bbox_output;
+        let (confidence_shape, confidence_data) = confidence_output;
+        let (landmark_shape, landmark_data) = landmark_output;
+
+        tracing::trace!(
+            "Output shapes - bbox: {:?}, confidence: {:?}, landmark: {:?}",
+            bbox_shape,
+            confidence_shape,
+            landmark_shape
+        );
+
+        // Convert to ndarray format
+        let bbox_dims = bbox_shape.as_ref();
+        let confidence_dims = confidence_shape.as_ref();
+        let landmark_dims = landmark_shape.as_ref();
+
+        let bbox_array = ndarray::Array3::from_shape_vec(
+            (
+                bbox_dims[0] as usize,
+                bbox_dims[1] as usize,
+                bbox_dims[2] as usize,
+            ),
+            bbox_data.to_vec(),
+        )
+        .change_context(Error)
+        .attach_printable("Failed to create bbox ndarray")?;
+
+        let confidence_array = ndarray::Array3::from_shape_vec(
+            (
+                confidence_dims[0] as usize,
+                confidence_dims[1] as usize,
+                confidence_dims[2] as usize,
+            ),
+            confidence_data.to_vec(),
+        )
+        .change_context(Error)
+        .attach_printable("Failed to create confidence ndarray")?;
+
+        let landmark_array = ndarray::Array3::from_shape_vec(
+            (
+                landmark_dims[0] as usize,
+                landmark_dims[1] as usize,
+                landmark_dims[2] as usize,
+            ),
+            landmark_data.to_vec(),
+        )
+        .change_context(Error)
+        .attach_printable("Failed to create landmark ndarray")?;
+
+        Ok(FaceDetectionModelOutput {
+            bbox: bbox_array,
+            confidence: confidence_array,
+            landmark: landmark_array,
+        })
+    }
+}

src/faceembed.rs

@@ -0,0 +1,35 @@
+pub mod facenet;
+
+// Re-export common types and traits
+pub use facenet::FaceNetEmbedder;
+pub use facenet::{FaceEmbedding, FaceEmbeddingConfig, IntoEmbeddings};
+
+// Convenience type aliases for different backends
+pub use facenet::mnn::EmbeddingGenerator as MnnEmbeddingGenerator;
+pub use facenet::ort::EmbeddingGenerator as OrtEmbeddingGenerator;
+
+use crate::errors::*;
+use ndarray::{Array2, ArrayView4};
+
+pub mod preprocessing {
+    use ndarray::*;
+    pub fn preprocess(faces: ArrayView4<u8>) -> Array4<f32> {
+        let mut owned = faces.as_standard_layout().mapv(|v| v as f32).to_owned();
+        owned.axis_iter_mut(Axis(0)).for_each(|mut image| {
+            let mean = image.mean().unwrap_or(0.0);
+            let std = image.std(0.0);
+            if std > 0.0 {
+                image.mapv_inplace(|x| (x - mean) / std);
+            } else {
+                image.mapv_inplace(|x| (x - 127.5) / 128.0)
+            }
+        });
+        owned
+    }
+}
+
+/// Common trait for face embedding backends - maintained for backward compatibility
+pub trait FaceEmbedder {
+    /// Generate embeddings for a batch of face images
+    fn run_models(&mut self, faces: ArrayView4<u8>) -> Result<Array2<f32>>;
+}

src/faceembed/facenet.rs

@@ -0,0 +1,209 @@
+pub mod mnn;
+pub mod ort;
+
+use crate::errors::*;
+use error_stack::ResultExt;
+use ndarray::{Array1, Array2, ArrayView3, ArrayView4};
+use ndarray_math::{CosineSimilarity, EuclideanDistance};
+
+/// Configuration for face embedding processing
+#[derive(Debug, Clone, PartialEq)]
+pub struct FaceEmbeddingConfig {
+    /// Input image width expected by the model
+    pub input_width: usize,
+    /// Input image height expected by the model
+    pub input_height: usize,
+    /// Whether to normalize embeddings to unit vectors
+    pub normalize: bool,
+}
+
+impl FaceEmbeddingConfig {
+    pub fn with_input_size(mut self, width: usize, height: usize) -> Self {
+        self.input_width = width;
+        self.input_height = height;
+        self
+    }
+
+    pub fn with_normalization(mut self, normalize: bool) -> Self {
+        self.normalize = normalize;
+        self
+    }
+}
+
+impl Default for FaceEmbeddingConfig {
+    fn default() -> Self {
+        Self {
+            input_width: 320,
+            input_height: 320,
+            normalize: false,
+        }
+    }
+}
+
+/// Represents a face embedding vector
+#[derive(Debug, Clone, PartialEq)]
+pub struct FaceEmbedding {
+    /// The embedding vector
+    pub vector: Array1<f32>,
+    /// Optional confidence score for the embedding quality
+    pub confidence: Option<f32>,
+}
+
+impl FaceEmbedding {
+    pub fn new(vector: Array1<f32>) -> Self {
+        Self {
+            vector,
+            confidence: None,
+        }
+    }
+
+    pub fn with_confidence(mut self, confidence: f32) -> Self {
+        self.confidence = Some(confidence);
+        self
+    }
+
+    /// Calculate cosine similarity with another embedding
+    pub fn cosine_similarity(&self, other: &FaceEmbedding) -> f32 {
+        self.vector.cosine_similarity(&other.vector).unwrap_or(0.0)
+    }
+
+    /// Calculate Euclidean distance with another embedding
+    pub fn euclidean_distance(&self, other: &FaceEmbedding) -> f32 {
+        self.vector
+            .euclidean_distance(other.vector.view())
+            .unwrap_or(f32::INFINITY)
+    }
+
+    /// Normalize the embedding vector to unit length
+    pub fn normalize(&mut self) {
+        let norm = self.vector.mapv(|x| x * x).sum().sqrt();
+        if norm > 0.0 {
+            self.vector.mapv_inplace(|x| x / norm);
+        }
+    }
+
+    /// Get the dimensionality of the embedding
+    pub fn dimension(&self) -> usize {
+        self.vector.len()
+    }
+}
+
+/// Raw model outputs that can be converted to embeddings
+pub trait IntoEmbeddings {
+    fn into_embeddings(self, config: &FaceEmbeddingConfig) -> Result<Vec<FaceEmbedding>>;
+}
+
+impl IntoEmbeddings for Array2<f32> {
+    fn into_embeddings(self, config: &FaceEmbeddingConfig) -> Result<Vec<FaceEmbedding>> {
+        let mut embeddings = Vec::new();
+
+        for row in self.rows() {
+            let mut vector = row.to_owned();
+
+            if config.normalize {
+                let norm = vector.mapv(|x| x * x).sum().sqrt();
+                if norm > 0.0 {
+                    vector.mapv_inplace(|x| x / norm);
+                }
+            }
+
+            embeddings.push(FaceEmbedding::new(vector));
+        }
+
+        Ok(embeddings)
+    }
+}
+
+/// Common trait for face embedding backends
+pub trait FaceNetEmbedder {
+    /// Generate embeddings for a batch of face images
+    fn run_model(&mut self, faces: ArrayView4<u8>) -> Result<Array2<f32>>;
+
+    /// Generate embeddings with full pipeline including postprocessing
+    fn generate_embeddings(
+        &mut self,
+        faces: ArrayView4<u8>,
+        config: FaceEmbeddingConfig,
+    ) -> Result<Vec<FaceEmbedding>> {
+        let raw_output = self
+            .run_model(faces)
+            .change_context(Error)
+            .attach_printable("Failed to generate embeddings")?;
+
+        raw_output
+            .into_embeddings(&config)
+            .attach_printable("Failed to process embeddings")
+    }
+
+    /// Generate a single embedding from a single face image
+    fn generate_embedding(
+        &mut self,
+        face: ArrayView3<u8>,
+        config: FaceEmbeddingConfig,
+    ) -> Result<FaceEmbedding> {
+        // Add batch dimension
+        let face_batch = face.insert_axis(ndarray::Axis(0));
+        let embeddings = self.generate_embeddings(face_batch.view(), config)?;
+
+        embeddings
+            .into_iter()
+            .next()
+            .ok_or(Error)
+            .attach_printable("No embedding generated for input face")
+    }
+}
+
+/// Utility functions for embedding processing
+pub mod utils {
+    use super::*;
+
+    /// Compute pairwise cosine similarities between two sets of embeddings
+    pub fn pairwise_cosine_similarities(
+        embeddings1: &[FaceEmbedding],
+        embeddings2: &[FaceEmbedding],
+    ) -> Array2<f32> {
+        let n1 = embeddings1.len();
+        let n2 = embeddings2.len();
+        let mut similarities = Array2::zeros((n1, n2));
+
+        for (i, emb1) in embeddings1.iter().enumerate() {
+            for (j, emb2) in embeddings2.iter().enumerate() {
+                similarities[(i, j)] = emb1.cosine_similarity(emb2);
+            }
+        }
+
+        similarities
+    }
+
+    /// Find the best matching embedding from a gallery for each query
+    pub fn find_best_matches(
+        queries: &[FaceEmbedding],
+        gallery: &[FaceEmbedding],
+    ) -> Vec<(usize, f32)> {
+        let similarities = pairwise_cosine_similarities(queries, gallery);
+        let mut best_matches = Vec::new();
+
+        for i in 0..queries.len() {
+            let row = similarities.row(i);
+            let (best_idx, best_score) = row
+                .iter()
+                .enumerate()
+                .max_by(|(_, a), (_, b)| a.partial_cmp(b).unwrap())
+                .unwrap();
+            best_matches.push((best_idx, *best_score));
+        }
+
+        best_matches
+    }
+
+    /// Filter embeddings by minimum quality threshold
+    pub fn filter_by_confidence(
+        embeddings: Vec<FaceEmbedding>,
+        min_confidence: f32,
+    ) -> Vec<FaceEmbedding> {
+        embeddings
+            .into_iter()
+            .filter(|emb| emb.confidence.map_or(true, |conf| conf >= min_confidence))
+            .collect()
+    }
+}

src/faceembed/facenet/mnn.rs

@@ -0,0 +1,127 @@
+use crate::errors::*;
+use crate::faceembed::facenet::FaceNetEmbedder;
+use mnn_bridge::ndarray::*;
+use ndarray::{Array1, Array2, ArrayView3, ArrayView4};
+use std::path::Path;
+
+#[derive(Debug)]
+pub struct EmbeddingGenerator {
+    handle: mnn_sync::SessionHandle,
+}
+pub struct EmbeddingGeneratorBuilder {
+    schedule_config: Option<mnn::ScheduleConfig>,
+    backend_config: Option<mnn::BackendConfig>,
+    model: mnn::Interpreter,
+}
+
+impl EmbeddingGeneratorBuilder {
+    pub fn new(model: impl AsRef<[u8]>) -> Result<Self> {
+        Ok(Self {
+            schedule_config: None,
+            backend_config: None,
+            model: mnn::Interpreter::from_bytes(model.as_ref())
+                .map_err(|e| e.into_inner())
+                .change_context(Error)
+                .attach_printable("Failed to load model from bytes")?,
+        })
+    }
+
+    pub fn with_forward_type(mut self, forward_type: mnn::ForwardType) -> Self {
+        self.schedule_config
+            .get_or_insert_default()
+            .set_type(forward_type);
+        self
+    }
+
+    pub fn with_schedule_config(mut self, config: mnn::ScheduleConfig) -> Self {
+        self.schedule_config = Some(config);
+        self
+    }
+
+    pub fn with_backend_config(mut self, config: mnn::BackendConfig) -> Self {
+        self.backend_config = Some(config);
+        self
+    }
+
+    pub fn build(self) -> Result<EmbeddingGenerator> {
+        let model = self.model;
+        let sc = self.schedule_config.unwrap_or_default();
+        let handle = mnn_sync::SessionHandle::new(model, sc)
+            .change_context(Error)
+            .attach_printable("Failed to create session handle")?;
+        Ok(EmbeddingGenerator { handle })
+    }
+}
+
+impl EmbeddingGenerator {
+    const INPUT_NAME: &'static str = "serving_default_input_6:0";
+    const OUTPUT_NAME: &'static str = "StatefulPartitionedCall:0";
+
+    pub fn builder<T: AsRef<[u8]>>(
+        model: T,
+    ) -> std::result::Result<EmbeddingGeneratorBuilder, Report<Error>> {
+        EmbeddingGeneratorBuilder::new(model)
+    }
+
+    pub fn run_models(&self, face: ArrayView4<u8>) -> Result<Array2<f32>> {
+        let tensor = crate::faceembed::preprocessing::preprocess(face);
+        let shape: [usize; 4] = tensor.dim().into();
+        let shape = shape.map(|f| f as i32);
+        let output = self
+            .handle
+            .run(move |sr| {
+                let tensor = tensor
+                    .as_mnn_tensor()
+                    .attach_printable("Failed to convert ndarray to mnn tensor")
+                    .change_context(mnn::ErrorKind::TensorError)?;
+                tracing::trace!("Image Tensor shape: {:?}", tensor.shape());
+                let (intptr, session) = sr.both_mut();
+                tracing::trace!("Copying input tensor to host");
+                let needs_resize = unsafe {
+                    let mut input = intptr.input_unresized::<f32>(session, Self::INPUT_NAME)?;
+                    tracing::trace!("Input shape: {:?}", input.shape());
+                    if *input.shape() != shape {
+                        tracing::trace!("Resizing input tensor to shape: {:?}", shape);
+                        // input.resize(shape);
+                        intptr.resize_tensor(input.view_mut(), shape);
+                        true
+                    } else {
+                        false
+                    }
+                };
+                if needs_resize {
+                    tracing::trace!("Resized input tensor to shape: {:?}", shape);
+                    let now = std::time::Instant::now();
+                    intptr.resize_session(session);
+                    tracing::trace!("Session resized in {:?}", now.elapsed());
+                }
+                let mut input = intptr.input::<f32>(session, Self::INPUT_NAME)?;
+                tracing::trace!("Input shape: {:?}", input.shape());
+                input.copy_from_host_tensor(tensor.view())?;
+
+                tracing::info!("Running face detection session");
+                intptr.run_session(&session)?;
+                let output_tensor = intptr
+                    .output::<f32>(&session, Self::OUTPUT_NAME)?
+                    .create_host_tensor_from_device(true)
+                    .as_ndarray()
+                    .to_owned();
+                Ok(output_tensor)
+            })
+            .change_context(Error)?;
+        Ok(output)
+    }
+}
+
+impl FaceNetEmbedder for EmbeddingGenerator {
+    fn run_model(&mut self, faces: ArrayView4<u8>) -> Result<Array2<f32>> {
+        self.run_models(faces)
+    }
+}
+
+// Main trait implementation for backward compatibility
+impl crate::faceembed::FaceEmbedder for EmbeddingGenerator {
+    fn run_models(&mut self, faces: ArrayView4<u8>) -> Result<Array2<f32>> {
+        EmbeddingGenerator::run_models(self, faces)
+    }
+}

src/faceembed/facenet/ort.rs

@@ -0,0 +1,207 @@
+use crate::errors::*;
+use crate::faceembed::facenet::FaceNetEmbedder;
+use crate::ort_ep::*;
+use error_stack::ResultExt;
+use ndarray::{Array2, ArrayView4};
+use ort::{
+    execution_providers::{CPUExecutionProvider, ExecutionProviderDispatch},
+    session::{Session, builder::GraphOptimizationLevel},
+    value::Tensor,
+};
+use std::path::Path;
+
+#[derive(Debug)]
+pub struct EmbeddingGenerator {
+    session: Session,
+}
+
+pub struct EmbeddingGeneratorBuilder {
+    model_data: Vec<u8>,
+    execution_providers: Option<Vec<ExecutionProviderDispatch>>,
+    intra_threads: Option<usize>,
+    inter_threads: Option<usize>,
+}
+
+impl EmbeddingGeneratorBuilder {
+    pub fn new(model: impl AsRef<[u8]>) -> crate::errors::Result<Self> {
+        Ok(Self {
+            model_data: model.as_ref().to_vec(),
+            execution_providers: None,
+            intra_threads: None,
+            inter_threads: None,
+        })
+    }
+
+    pub fn with_execution_providers(mut self, providers: impl AsRef<[ExecutionProvider]>) -> Self {
+        let execution_providers: Vec<ExecutionProviderDispatch> = providers
+            .as_ref()
+            .iter()
+            .filter_map(|provider| provider.to_dispatch())
+            .collect();
+
+        if !execution_providers.is_empty() {
+            self.execution_providers = Some(execution_providers);
+        } else {
+            tracing::warn!("No valid execution providers found, falling back to CPU");
+            self.execution_providers = Some(vec![CPUExecutionProvider::default().build()]);
+        }
+        self
+    }
+
+    pub fn with_intra_threads(mut self, threads: usize) -> Self {
+        self.intra_threads = Some(threads);
+        self
+    }
+
+    pub fn with_inter_threads(mut self, threads: usize) -> Self {
+        self.inter_threads = Some(threads);
+        self
+    }
+
+    pub fn build(self) -> crate::errors::Result<EmbeddingGenerator> {
+        let mut session_builder = Session::builder()
+            .change_context(Error)
+            .attach_printable("Failed to create session builder")?;
+
+        // Set execution providers
+        if let Some(providers) = self.execution_providers {
+            session_builder = session_builder
+                .with_execution_providers(providers)
+                .change_context(Error)
+                .attach_printable("Failed to set execution providers")?;
+        } else {
+            // Default to CPU
+            session_builder = session_builder
+                .with_execution_providers([CPUExecutionProvider::default().build()])
+                .change_context(Error)
+                .attach_printable("Failed to set default CPU execution provider")?;
+        }
+
+        // Set threading options
+        if let Some(threads) = self.intra_threads {
+            session_builder = session_builder
+                .with_intra_threads(threads)
+                .change_context(Error)
+                .attach_printable("Failed to set intra threads")?;
+        }
+
+        if let Some(threads) = self.inter_threads {
+            session_builder = session_builder
+                .with_inter_threads(threads)
+                .change_context(Error)
+                .attach_printable("Failed to set inter threads")?;
+        }
+
+        // Set optimization level
+        session_builder = session_builder
+            .with_optimization_level(GraphOptimizationLevel::Level3)
+            .change_context(Error)
+            .attach_printable("Failed to set optimization level")?;
+
+        // Create session from model bytes
+        let session = session_builder
+            .commit_from_memory(&self.model_data)
+            .change_context(Error)
+            .attach_printable("Failed to create ORT session from model bytes")?;
+
+        tracing::info!("Successfully created ORT FaceNet session");
+
+        Ok(EmbeddingGenerator { session })
+    }
+}
+
+impl EmbeddingGenerator {
+    const INPUT_NAME: &'static str = "serving_default_input_6:0";
+    const OUTPUT_NAME: &'static str = "StatefulPartitionedCall:0";
+
+    pub fn builder<T: AsRef<[u8]>>(
+        model: T,
+    ) -> std::result::Result<EmbeddingGeneratorBuilder, error_stack::Report<crate::errors::Error>>
+    {
+        EmbeddingGeneratorBuilder::new(model)
+    }
+
+    pub fn new(path: impl AsRef<Path>) -> crate::errors::Result<Self> {
+        let model = std::fs::read(path)
+            .change_context(Error)
+            .attach_printable("Failed to read model file")?;
+        Self::new_from_bytes(&model)
+    }
+
+    pub fn new_from_bytes(model: impl AsRef<[u8]>) -> crate::errors::Result<Self> {
+        tracing::info!("Loading ORT face embedding model from bytes");
+        Self::builder(model)?.build()
+    }
+
+    pub fn run_models(&mut self, faces: ArrayView4<u8>) -> crate::errors::Result<Array2<f32>> {
+        // Convert input from u8 to f32 and normalize to [0, 1] range
+        let input_tensor = crate::faceembed::preprocessing::preprocess(faces);
+
+        // face_array = np.asarray(face_resized, 'float32')
+        //  mean, std = face_array.mean(), face_array.std()
+        //  face_normalized = (face_array - mean) / std
+        // let input_tensor = faces.mean()
+
+        tracing::trace!("Input tensor shape: {:?}", input_tensor.shape());
+
+        // Create ORT input tensor
+        let input_value = Tensor::from_array(input_tensor)
+            .change_context(Error)
+            .attach_printable("Failed to create input tensor")?;
+
+        // Run inference
+        tracing::debug!("Running ORT FaceNet inference");
+        let outputs = self
+            .session
+            .run(ort::inputs![Self::INPUT_NAME => input_value])
+            .change_context(Error)
+            .attach_printable("Failed to run inference")?;
+
+        // Extract output tensor
+        let output = outputs
+            .get(Self::OUTPUT_NAME)
+            .ok_or(Error)
+            .attach_printable("Missing output from FaceNet model")?
+            .try_extract_tensor::<f32>()
+            .change_context(Error)
+            .attach_printable("Failed to extract output tensor")?;
+
+        let (output_shape, output_data) = output;
+
+        tracing::trace!("Output shape: {:?}", output_shape);
+
+        // Convert to ndarray format
+        let output_dims = output_shape.as_ref();
+
+        // FaceNet typically outputs embeddings as [batch_size, embedding_dim]
+        let batch_size = output_dims[0] as usize;
+        let embedding_dim = output_dims[1] as usize;
+
+        let output_array =
+            ndarray::Array2::from_shape_vec((batch_size, embedding_dim), output_data.to_vec())
+                .change_context(Error)
+                .attach_printable("Failed to create output ndarray")?;
+
+        tracing::trace!(
+            "Generated embeddings with shape: {:?}",
+            output_array.shape()
+        );
+
+        Ok(output_array)
+    }
+}
+
+impl FaceNetEmbedder for EmbeddingGenerator {
+    fn run_model(&mut self, faces: ArrayView4<u8>) -> crate::errors::Result<Array2<f32>> {
+        self.run_models(faces)
+    }
+}
+
+// Main trait implementation for backward compatibility
+impl crate::faceembed::FaceEmbedder for EmbeddingGenerator {
+    fn run_models(&mut self, faces: ArrayView4<u8>) -> crate::errors::Result<Array2<f32>> {
+        // Need to create a mutable reference for the session
+        // This is a workaround for the trait signature mismatch
+        self.run_models(faces)
+    }
+}

src/gui/bridge.rs

@@ -0,0 +1,569 @@
+use std::path::PathBuf;
+
+use crate::errors;
+use crate::facedet::{FaceDetectionConfig, FaceDetector, retinaface};
+use crate::faceembed::facenet;
+use crate::gui::app::{ComparisonResult, DetectionResult, ExecutorType};
+use bounding_box::Aabb2;
+use bounding_box::roi::MultiRoi as _;
+use error_stack::ResultExt;
+use fast_image_resize::ResizeOptions;
+use ndarray::{Array1, Array2, Array3, Array4};
+use ndarray_image::ImageToNdarray;
+use ndarray_math::CosineSimilarity;
+use ndarray_resize::NdFir;
+
+const RETINAFACE_MODEL_MNN: &[u8] = include_bytes!("../../models/retinaface.mnn");
+const FACENET_MODEL_MNN: &[u8] = include_bytes!("../../models/facenet.mnn");
+const RETINAFACE_MODEL_ONNX: &[u8] = include_bytes!("../../models/retinaface.onnx");
+const FACENET_MODEL_ONNX: &[u8] = include_bytes!("../../models/facenet.onnx");
+
+pub struct FaceDetectionBridge;
+
+impl FaceDetectionBridge {
+    pub async fn detect_faces(
+        image_path: PathBuf,
+        output_path: Option<PathBuf>,
+        threshold: f32,
+        nms_threshold: f32,
+        executor_type: ExecutorType,
+    ) -> DetectionResult {
+        let start_time = std::time::Instant::now();
+
+        match Self::run_detection_internal(
+            image_path.clone(),
+            output_path,
+            threshold,
+            nms_threshold,
+            executor_type,
+        )
+        .await
+        {
+            Ok((faces_count, processed_image)) => {
+                let processing_time = start_time.elapsed().as_secs_f64();
+                DetectionResult::Success {
+                    image_path,
+                    faces_count,
+                    processed_image,
+                    processing_time,
+                }
+            }
+            Err(error) => DetectionResult::Error(error.to_string()),
+        }
+    }
+
+    pub async fn compare_faces(
+        image1_path: PathBuf,
+        image2_path: PathBuf,
+        threshold: f32,
+        nms_threshold: f32,
+        executor_type: ExecutorType,
+    ) -> ComparisonResult {
+        let start_time = std::time::Instant::now();
+
+        match Self::run_comparison_internal(
+            image1_path,
+            image2_path,
+            threshold,
+            nms_threshold,
+            executor_type,
+        )
+        .await
+        {
+            Ok((
+                image1_faces,
+                image2_faces,
+                image1_face_rois,
+                image2_face_rois,
+                best_similarity,
+            )) => {
+                let processing_time = start_time.elapsed().as_secs_f64();
+                ComparisonResult::Success {
+                    image1_faces,
+                    image2_faces,
+                    image1_face_rois,
+                    image2_face_rois,
+                    best_similarity,
+                    processing_time,
+                }
+            }
+            Err(error) => ComparisonResult::Error(error.to_string()),
+        }
+    }
+
+    async fn run_detection_internal(
+        image_path: PathBuf,
+        output_path: Option<PathBuf>,
+        threshold: f32,
+        nms_threshold: f32,
+        executor_type: ExecutorType,
+    ) -> Result<(usize, Option<Vec<u8>>), Box<dyn std::error::Error + Send + Sync>> {
+        // Load the image
+        let img = image::open(&image_path)?;
+        let img_rgb = img.to_rgb8();
+
+        // Convert to ndarray format
+        let image_array = img_rgb.as_ndarray()?;
+
+        // Create detection configuration
+        let config = FaceDetectionConfig::default()
+            .with_threshold(threshold)
+            .with_nms_threshold(nms_threshold)
+            .with_input_width(1024)
+            .with_input_height(1024);
+
+        // Create detector and detect faces
+        let faces = match executor_type {
+            ExecutorType::MnnCpu => {
+                let mut detector = retinaface::mnn::FaceDetection::builder(RETINAFACE_MODEL_MNN)
+                    .map_err(|e| format!("Failed to create MNN detector: {}", e))?
+                    .with_forward_type(mnn::ForwardType::CPU)
+                    .build()
+                    .map_err(|e| format!("Failed to build MNN detector: {}", e))?;
+
+                detector
+                    .detect_faces(image_array.view(), &config)
+                    .map_err(|e| format!("Detection failed: {}", e))?
+            }
+            #[cfg(feature = "mnn-metal")]
+            ExecutorType::MnnMetal => {
+                let mut detector = retinaface::mnn::FaceDetection::builder(RETINAFACE_MODEL_MNN)
+                    .map_err(|e| format!("Failed to create MNN detector: {}", e))?
+                    .with_forward_type(mnn::ForwardType::Metal)
+                    .build()
+                    .map_err(|e| format!("Failed to build MNN detector: {}", e))?;
+
+                detector
+                    .detect_faces(image_array.view(), &config)
+                    .map_err(|e| format!("Detection failed: {}", e))?
+            }
+            #[cfg(feature = "mnn-coreml")]
+            ExecutorType::MnnCoreML => {
+                let mut detector = retinaface::mnn::FaceDetection::builder(RETINAFACE_MODEL_MNN)
+                    .map_err(|e| format!("Failed to create MNN detector: {}", e))?
+                    .with_forward_type(mnn::ForwardType::CoreML)
+                    .build()
+                    .map_err(|e| format!("Failed to build MNN detector: {}", e))?;
+
+                detector
+                    .detect_faces(image_array.view(), &config)
+                    .map_err(|e| format!("Detection failed: {}", e))?
+            }
+            ExecutorType::OnnxCpu => {
+                let mut detector = retinaface::ort::FaceDetection::builder(RETINAFACE_MODEL_ONNX)
+                    .map_err(|e| format!("Failed to create ONNX detector: {}", e))?
+                    .build()
+                    .map_err(|e| format!("Failed to build ONNX detector: {}", e))?;
+
+                detector
+                    .detect_faces(image_array.view(), &config)
+                    .map_err(|e| format!("Detection failed: {}", e))?
+            }
+            #[cfg(feature = "ort-cuda")]
+            ExecutorType::OrtCuda => {
+                use crate::ort_ep::ExecutionProvider;
+
+                let ep = ExecutionProvider::CUDA;
+                let mut detector = retinaface::ort::FaceDetection::builder(RETINAFACE_MODEL_ONNX)
+                    .map_err(|e| format!("Failed to create ONNX CUDA detector: {}", e))?
+                    .with_execution_providers([ep])
+                    .build()
+                    .map_err(|e| format!("Failed to build ONNX CUDA detector: {}", e))?;
+
+                detector
+                    .detect_faces(image_array.view(), &config)
+                    .map_err(|e| format!("CUDA detection failed: {}", e))?
+            }
+        };
+
+        let faces_count = faces.bbox.len();
+
+        // Generate output image with bounding boxes if requested
+        let processed_image = if output_path.is_some() || true {
+            // Always generate for GUI display
+            let mut output_img = img.to_rgb8();
+
+            for bbox in &faces.bbox {
+                let min_point = bbox.min_vertex();
+                let size = bbox.size();
+                let rect = imageproc::rect::Rect::at(min_point.x as i32, min_point.y as i32)
+                    .of_size(size.x as u32, size.y as u32);
+                imageproc::drawing::draw_hollow_rect_mut(
+                    &mut output_img,
+                    rect,
+                    image::Rgb([255, 0, 0]),
+                );
+            }
+
+            // Convert to bytes for GUI display
+            let mut buffer = Vec::new();
+            let mut cursor = std::io::Cursor::new(&mut buffer);
+            image::DynamicImage::ImageRgb8(output_img.clone())
+                .write_to(&mut cursor, image::ImageFormat::Png)?;
+
+            // Save to file if output path is specified
+            if let Some(ref output_path) = output_path {
+                output_img.save(output_path)?;
+            }
+
+            Some(buffer)
+        } else {
+            None
+        };
+
+        Ok((faces_count, processed_image))
+    }
+
+    async fn run_comparison_internal(
+        image1_path: PathBuf,
+        image2_path: PathBuf,
+        threshold: f32,
+        nms_threshold: f32,
+        executor_type: ExecutorType,
+    ) -> Result<
+        (usize, usize, Vec<Array3<u8>>, Vec<Array3<u8>>, f32),
+        Box<dyn std::error::Error + Send + Sync>,
+    > {
+        // Create detector and embedder, detect faces and generate embeddings
+        let (image1_faces, image2_faces, image1_rois, image2_rois, best_similarity) =
+            match executor_type {
+                ExecutorType::MnnCpu => {
+                    let mut detector =
+                        retinaface::mnn::FaceDetection::builder(RETINAFACE_MODEL_MNN)
+                            .map_err(|e| format!("Failed to create MNN detector: {}", e))?
+                            .with_forward_type(mnn::ForwardType::CPU)
+                            .build()
+                            .map_err(|e| format!("Failed to build MNN detector: {}", e))?;
+
+                    let mut embedder = facenet::mnn::EmbeddingGenerator::builder(FACENET_MODEL_MNN)
+                        .map_err(|e| format!("Failed to create MNN embedder: {}", e))?
+                        .with_forward_type(mnn::ForwardType::CPU)
+                        .build()
+                        .map_err(|e| format!("Failed to build MNN embedder: {}", e))?;
+
+                    let img_1 = run_detection(
+                        image1_path,
+                        &mut detector,
+                        &mut embedder,
+                        threshold,
+                        nms_threshold,
+                        2,
+                    )?;
+                    let img_2 = run_detection(
+                        image2_path,
+                        &mut detector,
+                        &mut embedder,
+                        threshold,
+                        nms_threshold,
+                        2,
+                    )?;
+
+                    let image1_rois = img_1.rois;
+                    let image2_rois = img_2.rois;
+                    let image1_bbox_len = img_1.bbox.len();
+                    let image2_bbox_len = img_2.bbox.len();
+                    let best_similarity = compare_faces(&img_1.embeddings, &img_2.embeddings)?;
+
+                    (
+                        image1_bbox_len,
+                        image2_bbox_len,
+                        image1_rois,
+                        image2_rois,
+                        best_similarity,
+                    )
+                }
+                #[cfg(feature = "mnn-metal")]
+                ExecutorType::MnnMetal => {
+                    let mut detector =
+                        retinaface::mnn::FaceDetection::builder(RETINAFACE_MODEL_MNN)
+                            .map_err(|e| format!("Failed to create MNN detector: {}", e))?
+                            .with_forward_type(mnn::ForwardType::Metal)
+                            .build()
+                            .map_err(|e| format!("Failed to build MNN detector: {}", e))?;
+
+                    let mut embedder = facenet::mnn::EmbeddingGenerator::builder(FACENET_MODEL_MNN)
+                        .map_err(|e| format!("Failed to create MNN embedder: {}", e))?
+                        .with_forward_type(mnn::ForwardType::Metal)
+                        .build()
+                        .map_err(|e| format!("Failed to build MNN embedder: {}", e))?;
+
+                    let img_1 = run_detection(
+                        image1_path,
+                        &mut detector,
+                        &mut embedder,
+                        threshold,
+                        nms_threshold,
+                        2,
+                    )?;
+                    let img_2 = run_detection(
+                        image2_path,
+                        &mut detector,
+                        &mut embedder,
+                        threshold,
+                        nms_threshold,
+                        2,
+                    )?;
+
+                    let image1_rois = img_1.rois;
+                    let image2_rois = img_2.rois;
+                    let image1_bbox_len = img_1.bbox.len();
+                    let image2_bbox_len = img_2.bbox.len();
+                    let best_similarity = compare_faces(&img_1.embeddings, &img_2.embeddings)?;
+
+                    (
+                        image1_bbox_len,
+                        image2_bbox_len,
+                        image1_rois,
+                        image2_rois,
+                        best_similarity,
+                    )
+                }
+                #[cfg(feature = "mnn-coreml")]
+                ExecutorType::MnnCoreML => {
+                    let mut detector =
+                        retinaface::mnn::FaceDetection::builder(RETINAFACE_MODEL_MNN)
+                            .map_err(|e| format!("Failed to create MNN detector: {}", e))?
+                            .with_forward_type(mnn::ForwardType::CoreML)
+                            .build()
+                            .map_err(|e| format!("Failed to build MNN detector: {}", e))?;
+
+                    let mut embedder = facenet::mnn::EmbeddingGenerator::builder(FACENET_MODEL_MNN)
+                        .map_err(|e| format!("Failed to create MNN embedder: {}", e))?
+                        .with_forward_type(mnn::ForwardType::CoreML)
+                        .build()
+                        .map_err(|e| format!("Failed to build MNN embedder: {}", e))?;
+
+                    let img_1 = run_detection(
+                        image1_path,
+                        &mut detector,
+                        &mut embedder,
+                        threshold,
+                        nms_threshold,
+                        2,
+                    )?;
+                    let img_2 = run_detection(
+                        image2_path,
+                        &mut detector,
+                        &mut embedder,
+                        threshold,
+                        nms_threshold,
+                        2,
+                    )?;
+
+                    let image1_rois = img_1.rois;
+                    let image2_rois = img_2.rois;
+                    let image1_bbox_len = img_1.bbox.len();
+                    let image2_bbox_len = img_2.bbox.len();
+                    let best_similarity = compare_faces(&img_1.embeddings, &img_2.embeddings)?;
+
+                    (
+                        image1_bbox_len,
+                        image2_bbox_len,
+                        image1_rois,
+                        image2_rois,
+                        best_similarity,
+                    )
+                }
+                ExecutorType::OnnxCpu => unimplemented!("ONNX face comparison not yet implemented"),
+                #[cfg(feature = "ort-cuda")]
+                ExecutorType::OrtCuda => {
+                    use crate::ort_ep::ExecutionProvider;
+                    let ep = ExecutionProvider::CUDA;
+                    let mut detector =
+                        retinaface::ort::FaceDetection::builder(RETINAFACE_MODEL_ONNX)
+                            .map_err(|e| format!("Failed to create MNN detector: {}", e))?
+                            .with_execution_providers([ep])
+                            .build()
+                            .map_err(|e| format!("Failed to build MNN detector: {}", e))?;
+
+                    let mut embedder =
+                        facenet::ort::EmbeddingGenerator::builder(FACENET_MODEL_ONNX)
+                            .map_err(|e| format!("Failed to create MNN embedder: {}", e))?
+                            .with_execution_providers([ep])
+                            .build()
+                            .map_err(|e| format!("Failed to build MNN embedder: {}", e))?;
+
+                    let img_1 = run_detection(
+                        image1_path,
+                        &mut detector,
+                        &mut embedder,
+                        threshold,
+                        nms_threshold,
+                        2,
+                    )?;
+                    let img_2 = run_detection(
+                        image2_path,
+                        &mut detector,
+                        &mut embedder,
+                        threshold,
+                        nms_threshold,
+                        2,
+                    )?;
+
+                    let image1_rois = img_1.rois;
+                    let image2_rois = img_2.rois;
+                    let image1_bbox_len = img_1.bbox.len();
+                    let image2_bbox_len = img_2.bbox.len();
+                    let best_similarity = compare_faces(&img_1.embeddings, &img_2.embeddings)?;
+
+                    (
+                        image1_bbox_len,
+                        image2_bbox_len,
+                        image1_rois,
+                        image2_rois,
+                        best_similarity,
+                    )
+                }
+            };
+
+        Ok((
+            image1_faces,
+            image2_faces,
+            image1_rois,
+            image2_rois,
+            best_similarity,
+        ))
+    }
+}
+
+use crate::errors::Error;
+pub fn compare_faces(
+    faces_1: &[Array1<f32>],
+    faces_2: &[Array1<f32>],
+) -> Result<f32, error_stack::Report<crate::errors::Error>> {
+    use error_stack::Report;
+
+    if faces_1.is_empty() || faces_2.is_empty() {
+        Err(Report::new(crate::errors::Error))
+            .attach_printable("One or both images have no detected faces")?;
+    }
+    if faces_1.len() != faces_2.len() {
+        Err(Report::new(crate::errors::Error))
+            .attach_printable("Face count mismatch between images")?;
+    }
+    Ok(faces_1
+        .iter()
+        .zip(faces_2)
+        .flat_map(|(face_1, face_2)| face_1.cosine_similarity(face_2))
+        .inspect(|v| tracing::info!("Cosine similarity: {}", v))
+        .map(|v| ordered_float::OrderedFloat(v))
+        .max()
+        .map(|v| v.0)
+        .ok_or(Report::new(Error))?)
+}
+
+#[derive(Debug)]
+pub struct DetectionOutput {
+    bbox: Vec<Aabb2<usize>>,
+    rois: Vec<ndarray::Array3<u8>>,
+    embeddings: Vec<Array1<f32>>,
+}
+
+fn run_detection<D, E>(
+    image: impl AsRef<std::path::Path>,
+    retinaface: &mut D,
+    facenet: &mut E,
+    threshold: f32,
+    nms_threshold: f32,
+    chunk_size: usize,
+) -> crate::errors::Result<DetectionOutput>
+where
+    D: crate::facedet::FaceDetector,
+    E: crate::faceembed::FaceEmbedder,
+{
+    use errors::*;
+    // Initialize database if requested
+    let image = image.as_ref();
+    let image = image::open(image)
+        .change_context(Error)
+        .attach_printable(image.to_string_lossy().to_string())?;
+    let image = image.into_rgb8();
+    let mut array = image
+        .into_ndarray()
+        .change_context(errors::Error)
+        .attach_printable("Failed to convert image to ndarray")?;
+    let output = retinaface
+        .detect_faces(
+            array.view(),
+            &FaceDetectionConfig::default()
+                .with_threshold(threshold)
+                .with_nms_threshold(nms_threshold),
+        )
+        .change_context(errors::Error)
+        .attach_printable("Failed to detect faces")?;
+
+    let bboxes = output
+        .bbox
+        .iter()
+        .inspect(|bbox| tracing::info!("Raw bbox: {:?}", bbox))
+        .map(|bbox| bbox.as_::<f32>().scale_uniform(1.30).as_::<usize>())
+        .inspect(|bbox| tracing::info!("Padded bbox: {:?}", bbox))
+        .collect_vec();
+    for bbox in &bboxes {
+        tracing::info!("Detected face: {:?}", bbox);
+        use bounding_box::draw::*;
+        array.draw(bbox, color::palette::css::GREEN_YELLOW.to_rgba8(), 1);
+    }
+    use itertools::Itertools;
+    let face_rois = array
+        .view()
+        .multi_roi(&bboxes)
+        .change_context(Error)?
+        .into_iter()
+        .map(|roi| {
+            roi.as_standard_layout()
+                .fast_resize(224, 224, &ResizeOptions::default())
+                .change_context(Error)
+        })
+        .collect::<Result<Vec<_>>>()?;
+    let face_roi_views = face_rois.iter().map(|roi| roi.view()).collect::<Vec<_>>();
+
+    let embeddings: Vec<Array1<f32>> = face_roi_views
+        .chunks(chunk_size)
+        .map(|chunk| {
+            tracing::info!("Processing chunk of size: {}", chunk.len());
+
+            let og_size = chunk.len();
+            if chunk.len() < chunk_size {
+                tracing::warn!("Chunk size is less than 8, padding with zeros");
+                let zeros = Array3::zeros((224, 224, 3));
+                let chunk: Vec<_> = chunk
+                    .iter()
+                    .map(|arr| arr.reborrow())
+                    .chain(core::iter::repeat(zeros.view()))
+                    .take(chunk_size)
+                    .collect();
+                let face_rois: Array4<u8> = ndarray::stack(ndarray::Axis(0), chunk.as_slice())
+                    .change_context(errors::Error)
+                    .attach_printable("Failed to stack rois together")?;
+                let output = facenet.run_models(face_rois.view()).change_context(Error)?;
+                Ok((output, og_size))
+            } else {
+                let face_rois: Array4<u8> = ndarray::stack(ndarray::Axis(0), chunk)
+                    .change_context(errors::Error)
+                    .attach_printable("Failed to stack rois together")?;
+                let output = facenet.run_models(face_rois.view()).change_context(Error)?;
+                Ok((output, og_size))
+            }
+        })
+        .collect::<Result<Vec<(Array2<f32>, usize)>>>()?
+        .into_iter()
+        .map(|(chunk, size): (Array2<f32>, usize)| {
+            use itertools::Itertools;
+            chunk
+                .rows()
+                .into_iter()
+                .take(size)
+                .map(|row| row.to_owned())
+                .collect_vec()
+                .into_iter()
+        })
+        .flatten()
+        .collect::<Vec<Array1<f32>>>();
+
+    Ok(DetectionOutput {
+        bbox: bboxes,
+        rois: face_rois,
+        embeddings,
+    })
+}

src/gui/mod.rs

@@ -0,0 +1,5 @@
+pub mod app;
+pub mod bridge;
+
+pub use app::{FaceDetectorApp, Message, run};
+pub use bridge::FaceDetectionBridge;

src/image.rs

@@ -1,5 +0,0 @@
-// pub struct Image {
-//     pub width: u32,
-//     pub height: u32,
-//     pub data: Vec<u8>,
-// }

src/lib.rs

@@ -1,4 +1,7 @@
+pub mod database;
 pub mod errors;
 pub mod facedet;
-pub mod image;
-use errors::*;
+pub mod faceembed;
+pub mod gui;
+pub mod ort_ep;
+pub use errors::*;

src/main.rs

@@ -1,59 +0,0 @@
-mod cli;
-mod errors;
-use detector::facedet::retinaface::FaceDetectionConfig;
-use errors::*;
-use ndarray_image::*;
-const RETINAFACE_MODEL: &[u8] = include_bytes!("../models/retinaface.mnn");
-pub fn main() -> Result<()> {
-    tracing_subscriber::fmt()
-        .with_env_filter("trace")
-        .with_thread_ids(true)
-        .with_thread_names(true)
-        .with_target(false)
-        .init();
-    let args = <cli::Cli as clap::Parser>::parse();
-    match args.cmd {
-        cli::SubCommand::Detect(detect) => {
-            use detector::facedet;
-            let model = facedet::retinaface::FaceDetection::new_from_bytes(RETINAFACE_MODEL)
-                .change_context(errors::Error)
-                .attach_printable("Failed to create face detection model")?;
-            let image = image::open(detect.image).change_context(Error)?;
-            let image = image.into_rgb8();
-            let mut array = image
-                .into_ndarray()
-                .change_context(errors::Error)
-                .attach_printable("Failed to convert image to ndarray")?;
-            let output = model
-                .detect_faces(
-                    array.clone(),
-                    FaceDetectionConfig::default().with_threshold(detect.threshold),
-                )
-                .change_context(errors::Error)
-                .attach_printable("Failed to detect faces")?;
-            for bbox in output.bbox {
-                tracing::info!("Detected face: {:?}", bbox);
-                use bounding_box::draw::*;
-                array.draw(bbox, color::palette::css::GREEN_YELLOW.to_rgba8(), 1);
-            }
-            let v = array.view();
-            if let Some(output) = detect.output {
-                let image: image::RgbImage = v
-                    .to_image()
-                    .change_context(errors::Error)
-                    .attach_printable("Failed to convert ndarray to image")?;
-                image
-                    .save(output)
-                    .change_context(errors::Error)
-                    .attach_printable("Failed to save output image")?;
-            }
-        }
-        cli::SubCommand::List(list) => {
-            println!("List: {:?}", list);
-        }
-        cli::SubCommand::Completions { shell } => {
-            cli::Cli::completions(shell);
-        }
-    }
-    Ok(())
-}

src/ort_ep.rs

@@ -0,0 +1,197 @@
+#[cfg(feature = "ort-cuda")]
+use ort::execution_providers::CUDAExecutionProvider;
+#[cfg(feature = "ort-coreml")]
+use ort::execution_providers::CoreMLExecutionProvider;
+#[cfg(feature = "ort-directml")]
+use ort::execution_providers::DirectMLExecutionProvider;
+#[cfg(feature = "ort-openvino")]
+use ort::execution_providers::OpenVINOExecutionProvider;
+#[cfg(feature = "ort-tvm")]
+use ort::execution_providers::TVMExecutionProvider;
+#[cfg(feature = "ort-tensorrt")]
+use ort::execution_providers::TensorRTExecutionProvider;
+use ort::execution_providers::{CPUExecutionProvider, ExecutionProviderDispatch};
+
+/// Supported execution providers for ONNX Runtime
+#[derive(Debug, Copy, Clone)]
+pub enum ExecutionProvider {
+    /// CPU execution provider (always available)
+    CPU,
+    /// CoreML execution provider (macOS only)
+    CoreML,
+    /// CUDA execution provider (requires cuda feature)
+    CUDA,
+    /// TensorRT execution provider (requires tensorrt feature)
+    TensorRT,
+    /// TVM execution provider (requires tvm feature)
+    TVM,
+    /// OpenVINO execution provider (requires openvino feature)
+    OpenVINO,
+    /// DirectML execution provider (Windows only, requires directml feature)
+    DirectML,
+}
+
+impl std::fmt::Display for ExecutionProvider {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            ExecutionProvider::CPU => write!(f, "CPU"),
+            ExecutionProvider::CoreML => write!(f, "CoreML"),
+            ExecutionProvider::CUDA => write!(f, "CUDA"),
+            ExecutionProvider::TensorRT => write!(f, "TensorRT"),
+            ExecutionProvider::TVM => write!(f, "TVM"),
+            ExecutionProvider::OpenVINO => write!(f, "OpenVINO"),
+            ExecutionProvider::DirectML => write!(f, "DirectML"),
+        }
+    }
+}
+
+impl std::str::FromStr for ExecutionProvider {
+    type Err = String;
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s.to_lowercase().as_str() {
+            "cpu" => Ok(ExecutionProvider::CPU),
+            "coreml" => Ok(ExecutionProvider::CoreML),
+            "cuda" => Ok(ExecutionProvider::CUDA),
+            "tensorrt" => Ok(ExecutionProvider::TensorRT),
+            "tvm" => Ok(ExecutionProvider::TVM),
+            "openvino" => Ok(ExecutionProvider::OpenVINO),
+            "directml" => Ok(ExecutionProvider::DirectML),
+            _ => Err(format!("Unknown execution provider: {}", s)),
+        }
+    }
+}
+
+impl ExecutionProvider {
+    /// Returns all available execution providers for the current platform and features
+    pub fn available_providers() -> Vec<ExecutionProvider> {
+        vec![
+            ExecutionProvider::CPU,
+            #[cfg(all(target_os = "macos", feature = "ort-coreml"))]
+            ExecutionProvider::CoreML,
+            #[cfg(feature = "ort-cuda")]
+            ExecutionProvider::CUDA,
+            #[cfg(feature = "ort-tensorrt")]
+            ExecutionProvider::TensorRT,
+            #[cfg(feature = "ort-tvm")]
+            ExecutionProvider::TVM,
+            #[cfg(feature = "ort-openvino")]
+            ExecutionProvider::OpenVINO,
+            #[cfg(all(target_os = "windows", feature = "ort-directml"))]
+            ExecutionProvider::DirectML,
+        ]
+    }
+
+    /// Check if this execution provider is available on the current platform
+    pub fn is_available(&self) -> bool {
+        match self {
+            ExecutionProvider::CPU => true,
+            ExecutionProvider::CoreML => cfg!(target_os = "macos") && cfg!(feature = "ort-coreml"),
+            ExecutionProvider::CUDA => cfg!(feature = "ort-cuda"),
+            ExecutionProvider::TensorRT => cfg!(feature = "ort-tensorrt"),
+            ExecutionProvider::TVM => cfg!(feature = "ort-tvm"),
+            ExecutionProvider::OpenVINO => cfg!(feature = "ort-openvino"),
+            ExecutionProvider::DirectML => {
+                cfg!(target_os = "windows") && cfg!(feature = "ort-directml")
+            }
+        }
+    }
+}
+
+impl ExecutionProvider {
+    pub fn to_dispatch(&self) -> Option<ExecutionProviderDispatch> {
+        match self {
+            ExecutionProvider::CPU => Some(CPUExecutionProvider::default().build()),
+            ExecutionProvider::CoreML => {
+                #[cfg(target_os = "macos")]
+                {
+                    #[cfg(feature = "ort-coreml")]
+                    {
+                        use tap::Tap;
+
+                        Some(
+                            CoreMLExecutionProvider::default()
+                                .with_model_format(
+                                    ort::execution_providers::coreml::CoreMLModelFormat::MLProgram,
+                                )
+                                .build(),
+                        )
+                    }
+                    #[cfg(not(feature = "ort-coreml"))]
+                    {
+                        tracing::error!("coreml support not compiled in");
+                        None
+                    }
+                }
+                #[cfg(not(target_os = "macos"))]
+                {
+                    tracing::error!("CoreML is only available on macOS");
+                    None
+                }
+            }
+            ExecutionProvider::CUDA => {
+                #[cfg(feature = "ort-cuda")]
+                {
+                    Some(CUDAExecutionProvider::default().build())
+                }
+                #[cfg(not(feature = "ort-cuda"))]
+                {
+                    tracing::error!("CUDA support not compiled in");
+                    None
+                }
+            }
+            ExecutionProvider::TensorRT => {
+                #[cfg(feature = "ort-tensorrt")]
+                {
+                    Some(TensorRTExecutionProvider::default().build())
+                }
+                #[cfg(not(feature = "ort-tensorrt"))]
+                {
+                    tracing::error!("TensorRT support not compiled in");
+                    None
+                }
+            }
+            ExecutionProvider::TVM => {
+                #[cfg(feature = "ort-tvm")]
+                {
+                    Some(TVMExecutionProvider::default().build())
+                }
+                #[cfg(not(feature = "ort-tvm"))]
+                {
+                    tracing::error!("TVM support not compiled in");
+                    None
+                }
+            }
+            ExecutionProvider::OpenVINO => {
+                #[cfg(feature = "ort-openvino")]
+                {
+                    Some(OpenVINOExecutionProvider::default().build())
+                }
+                #[cfg(not(feature = "ort-openvino"))]
+                {
+                    tracing::error!("OpenVINO support not compiled in");
+                    None
+                }
+            }
+            ExecutionProvider::DirectML => {
+                #[cfg(target_os = "windows")]
+                {
+                    #[cfg(feature = "ort-directml")]
+                    {
+                        Some(DirectMLExecutionProvider::default().build())
+                    }
+                    #[cfg(not(feature = "ort-directml"))]
+                    {
+                        tracing::error!("DirectML support not compiled in");
+                        None
+                    }
+                }
+                #[cfg(not(target_os = "windows"))]
+                {
+                    tracing::error!("DirectML is only available on Windows");
+                    None
+                }
+            }
+        }
+    }
+}