open-body-tracker: obt-h3c
Single-RGB-camera full-body pose: 23 keypoints (COCO 17 body + 6 foot points) in 2D and root-relative metric 3D, from a 256ร192 person crop. Code, demos and documentation: https://github.com/yuru-lab/open-body-tracker ยท browser demo: https://yuru-lab.github.io/open-body-tracker/
| File | Format |
|---|---|
obt-h3c.safetensors |
PyTorch weights (EMA), load with open_body_tracker.weights.load_model |
obt-h3c.onnx |
ONNX fp32 (382 MB). Input crop float32 (N,3,256,192) RGB 0โ255; outputs xy (N,23,2) crop px, score (N,23), xyz (N,23,3) metres |
obt-h3c-fp16.onnx |
same, fp16 weights (191 MB, used by the WebGPU demo) |
yolox_nano.onnx, yolox_tiny.onnx |
Unmodified mirrors of the official Megvii YOLOX 0.1.1rc0 ONNX person/object detectors (Apache-2.0, YOLOX-LICENSE.txt), mirrored for browser CORS |
Results
| Benchmark | Metric | obt-h3c |
|---|---|---|
| COCO val2017 (2D, GT boxes, training-overlap images removed) | keypoint AP | 71.7 |
| ASPset-510, 3 held-out subjects | PA-MPJPE (body 12) | 67.9 mm |
| MPJPE | 92.7 mm |
Details, training data and limitations: MODEL_CARD.md.
License
The weights are distributed under the SAM License (LICENSE) because the model was trained in part on outputs
of SAM 3D Body (Meta). Include the license when redistributing; its use restrictions apply. Training images
(COCO 2014 and Open Images V7, CC BY 2.0 images only) are not included. YOLOX files: Apache-2.0.
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐ Ask for provider support