feat: output aabbs from retinaface

src/facedet/retinaface.rs

@@ -0,0 +1,213 @@
+use crate::errors::*;
+use bounding_box::Aabb2;
+use error_stack::ResultExt;
+use mnn_bridge::ndarray::*;
+use nalgebra::{Point2, Vector2};
+use ndarray_resize::NdFir;
+use std::path::Path;
+
+pub struct FaceDetectionConfig {
+    min_sizes: Vec<Vector2<usize>>,
+    steps: Vec<usize>,
+    variance: Vec<f32>,
+}
+
+impl Default for FaceDetectionConfig {
+    fn default() -> Self {
+        FaceDetectionConfig {
+            min_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],
+        }
+    }
+}
+pub struct FaceDetection {
+    handle: mnn_sync::SessionHandle,
+}
+
+pub struct FaceDetectionModelOutput {
+    pub bbox: ndarray::Array3<f32>,
+    pub confidence: ndarray::Array3<f32>,
+    pub landmark: ndarray::Array3<f32>,
+}
+
+impl FaceDetectionModelOutput {
+    pub fn postprocess(self, config: FaceDetectionConfig) -> Result<Vec<Aabb2<f32>>> {
+        let mut anchors = Vec::new();
+        for (k, &step) in config.steps.iter().enumerate() {
+            let feature_size = 640 / step;
+            let min_sizes = config.min_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 / 640.0;
+                        let cy = (i as f32 + 0.5) * step as f32 / 640.0;
+                        let s_k = size as f32 / 640.0;
+                        anchors.push((cx, cy, s_k, s_k));
+                    }
+                }
+            }
+        }
+        let mut boxes = Vec::new();
+        let var0 = config.variance[0];
+        let var1 = config.variance[1];
+        let bbox_data = self.bbox;
+        let conf_data = self.confidence;
+        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 > 0.6 {
+                boxes.push(Aabb2::from_min_max_vertices(
+                    Point2::new(x_min, y_min),
+                    Point2::new(x_max, y_max),
+                ));
+            }
+        }
+        Ok(boxes)
+    }
+}
+
+impl FaceDetectionModelOutput {
+    pub fn print(&self, limit: usize) {
+        tracing::info!("Detected {} faces", self.bbox.shape()[1]);
+
+        for (bbox, confidence) in self
+            .bbox
+            .clone()
+            .remove_axis(ndarray::Axis(0))
+            .axis_iter(ndarray::Axis(0))
+            .zip(
+                self.confidence
+                    .clone()
+                    .remove_axis(ndarray::Axis(0))
+                    .axis_iter(ndarray::Axis(0))
+                    .map(|c| c[1]),
+            )
+            .filter(|(_, c)| *c > 0.1)
+            .take(limit)
+        {
+            tracing::info!("Face BBox: {:?}, Confidence: {:.2}", bbox, confidence);
+        }
+    }
+}
+
+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>) -> Result<FaceDetectionModelOutput> {
+        #[rustfmt::skip]
+        use ::tap::*;
+        let output = self
+            .handle
+            .run(move |sr| {
+                let mut resized = image
+                    .fast_resize(640, 640, 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,
+                        640,
+                        640,
+                    );
+                }
+                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/yolo.rs

@@ -0,0 +1 @@
+