feat: add support for HRNet model

depthai_nodes/ml/parsers/__init__.py

@@ -11,6 +11,7 @@
 from .xfeat import XFeatParser
 from .yunet import YuNetParser
 from .age_gender import AgeGenderParser
+from .hrnet import HRNetParser
 
 __all__ = [
     "ImageOutputParser",
@@ -26,4 +27,5 @@
     "XFeatParser",
     "ThermalImageParser",
     "AgeGenderParser",
+    "HRNetParser",
 ]

depthai_nodes/ml/parsers/hrnet.py

@@ -0,0 +1,62 @@
+import depthai as dai
+import numpy as np
+import cv2
+
+from ..messages.creators import create_keypoints_message
+
+
+class HRNetParser(dai.node.ThreadedHostNode):
+    def __init__(self, score_threshold=0.5, input_size=[256, 256], heatmap_size=[64, 64]):
+        dai.node.ThreadedHostNode.__init__(self)
+        self.input = dai.Node.Input(self)
+        self.out = dai.Node.Output(self)
+        self.input_size = input_size
+        self.heatmap_size = heatmap_size
+        self.score_threshold = score_threshold
+
+    def setScoreThreshold(self, threshold):
+        self.score_threshold = threshold
+
+    def run(self):
+        """Postprocessing logic for HRNet pose estimation model. The code is inspired by https://github.com/ibaiGorordo/ONNX-HRNET-Human-Pose-Estimation
+        
+        Returns:
+            ...
+        """
+
+        while self.isRunning():
+            try:
+                output: dai.NNData = self.input.get()
+            except dai.MessageQueue.QueueException:
+                break  # Pipeline was stopped
+
+            img_width, img_height = self.input_size
+
+            heatmaps = output.getTensor("heatmaps", dequantize=True)
+
+            if len(heatmaps.shape) == 4: # add new axis for batch size
+                heatmaps = heatmaps[0]
+
+            if heatmaps.shape[2] == 16: # HW_ instead of _HW
+                heatmaps = heatmaps.transpose(2, 0, 1)
+
+            _, map_h, map_w = heatmaps.shape
+
+            # Find the maximum value in each of the heatmaps and its location
+            max_vals = np.array([np.max(heatmap) for heatmap in heatmaps])
+            keypoints = np.array([np.unravel_index(heatmap.argmax(), heatmap.shape)
+                                for heatmap in heatmaps])
+            keypoints = keypoints.astype(np.float32)
+            keypoints[max_vals < self.score_threshold] = np.array([np.nan, np.nan])
+
+            # Scale keypoints to the image size
+            # TODO: remove and have relative keypoint values? e.g. * np.array([64 / map_w, 64 / map_h]) to get relative values?
+            keypoints = keypoints[:, ::-1] * np.array([img_width / map_w, img_height / map_h])
+
+            keypoints_msg = create_keypoints_message(
+                keypoints=keypoints,
+                #scores=max_vals, # TODO: add scores
+                #confidence_threshold=self.confidence_threshold # TODO: add confidence threshold
+            )
+
+            self.out.send(keypoints_msg)