BlazePose person detector (ONNX)
Google MediaPipe's BlazePose detector, from the Pose Landmarker .task
bundle, converted to ONNX.
- Input
[1,224,224,3]RGB in[-1,1], letterboxed. - Outputs
[1,2254,12]+[1,2254,1]. Anchors: strides[8,16,32,32,32]→ 2254. The ROI is centred on the mid-hip keypoint, sized from the hip→scale keypoint distance, target angle 90°, 1.25× square.
The detector ships fp16 block-sparse weights behind DENSIFY ops that tf2onnx can't parse; the export densifies them via the TFLite interpreter first (see the export script).
License
Apache-2.0 — this graph is a direct ONNX conversion of a Google
MediaPipe model (Apache-2.0 code
AND weights). Conversion + numerical-parity proof (vs the Python mediapipe
reference): scripts/export-facecap-onnx.py,
contract in docs/MOCAP_SPIKE.md.
How it is used
Mirror of one graph from fernandotonon/QtMeshEditor-models
(mocap/…), which QtMeshEditor
downloads on first use for its Performance Capture feature (video/webcam →
facial morph + head + full-body skeletal animation, epic #869). This standalone
repo is for discoverability; the app fetches from the aggregate repo.