qtpose/synthpose-huge-52-onnx

SynthPose-Huge 52 keypoints (ONNX) for qtpose

ONNX export of stanfordmimi/synthpose-vitpose-huge-hf (ViTPose, about 637M parameters, fine-tuned from usyd-community/vitpose-plus-huge), published for use by qtpose. SynthPose predicts 52 anatomical landmarks, including spine, pelvis and foot points, for markerless motion capture.

Licence and citation

The SynthPose weights are released under the Apache-2.0 licence, and this derived ONNX file is distributed under the same licence. SynthPose is described in arXiv:2406.09788; please cite it when you use this model.

Export

Exported with qtpose's helper scripts/dev/convert_synthpose_to_onnx.py --variant huge --validate, which loads the model with transformers.VitPoseForPoseEstimation and runs torch.onnx.export(..., opset_version=17, dynamo=False) on a wrapper that returns only the heatmap tensor.

Model file

• synthpose_huge_52.zip, 2549609098 bytes: an uncompressed (stored) ZIP holding synthpose_huge_52.onnx and its ONNX external-data file(s); the model is larger than the 2 GB protobuf limit. qtpose extracts it after download.
• Input: pixel_values, float32, shape (batch, 3, 256, 192) (RGB person crop, ImageNet mean/std normalised)
• Output: heatmaps, float32, shape (batch, 52, 64, 48): one heatmap per keypoint at 1/4 of the input resolution. A person detector supplies the crop.
• SHA-256 of the downloaded file: 768ce11064893406db78245906df3eefc05494d7df35ad7ae875c2e314ca2f25