# RTMPose keypoints, COCO-WholeBody 133

Whole-body keypoints of every PJM segment (39,813 of the 39,817; the 4 clips that could not be opened are
listed in `crop_params/failed.csv`), as used to train the Uni-Sign models in `CrocoSign/pjm-models`.
Extracted with RTMPose (rtmlib `Wholebody`) and stored in the COCO-WholeBody layout of 133 points, the
layout the Uni-Sign code indexes.

Shards `pose_coco133-NN.tar` unpack to `pose_coco133/{key}.pkl`, one pickle per clip key (the same keys as
in `splits/`). Each pickle is a dict with `keypoints`, float64 of shape (frames, 1, 133, 2), and `scores`,
float64 of shape (frames, 1, 133).

```bash
for f in pose_coco133-*.tar; do tar -xf "$f"; done
```

Point Uni-Sign's `pose_dirs["PJM"]` (`config.py`) at the unpacked `pose_coco133/` directory.