OpenPose-135 weights (BODY_25 + hand + face)
Self-contained PyTorch checkpoints for the CMU OpenPose BODY_25 body (25 keypoints),
hand (21 keypoints × 2), and face (70 keypoints) detectors — the exact stack
needed to reproduce the OpenPose-135 keypoint layout without a Caffe runtime or mmcv.
Files
| File | Size | Source |
|---|---|---|
body_pose_model_25.pth |
~200 MB | CMU pose_iter_584000.caffemodel ported via caffemodel2pytorch, redistributed by TracelessLe/OpenPose.PyTorch |
hand_pose_model.pth |
147 MB | CMU hand pose model, mirrored from lllyasviel/Annotators |
facenet.pth |
154 MB | CMU OpenPose face / FaceNet, mirrored from lllyasviel/Annotators |
Usage
These are intended to be loaded by tools/openpose135/ — the
self-contained PyTorch OpenPose-135 detector vendored into the PHD codebase. The detector
auto-downloads from this repo on first use:
from tools.openpose135 import OpenPose135Detector
detector = OpenPose135Detector(device="cuda") # auto-fetches the three .pth files
people = detector(image_rgb)
To use these weights without that wrapper, see the architecture definitions in
tools/openpose135/model.py — the state dicts are flat (caffemodel2pytorch / direct
naming) and load through a small transfer() helper.
License
These weights are derived from the CMU OpenPose project, which is licensed for non-commercial use only:
The OpenPose project is freely available for free non-commercial use, and may be redistributed under these conditions. Please, see the LICENSE for further details. Interested in a commercial license? Contact the CMU Technology Transfer and Enterprise Creation office.
Redistribution here is non-commercial; downstream use inherits the same restriction.
Attribution
- CMU OpenPose: https://github.com/CMU-Perceptual-Computing-Lab/openpose
- BODY_25 PyTorch port: https://github.com/TracelessLe/OpenPose.PyTorch
- Hand + face PyTorch port: https://github.com/lllyasviel/ControlNet-v1-1-nightly
- Original pytorch-openpose: https://github.com/Hzzone/pytorch-openpose