# Third-party notice — MoveNet MultiPose Lightning

This directory contains an unmodified pretrained model checkpoint redistributed under its original license.

- **Model**: MoveNet MultiPose Lightning (up to 6 people, 17 COCO keypoints each)
- **Original model**: Google (MoveNet, released via TensorFlow Hub, Apache-2.0)
- **ONNX conversion**: Xenova (converted for browser / Transformers.js use)
- **Upstream repository**: https://huggingface.co/Xenova/movenet-multipose-lightning
- **License**: Apache License 2.0 (see `LICENSE` in this directory)
- **Commercial use**: Permitted. Apache-2.0 has no non-commercial restriction; it only requires retaining the license/copyright notice and stating any changes made to the source.
- **Modifications made**: None. `model.onnx` is used exactly as published upstream.
- **Asset source**: https://huggingface.co/Xenova/movenet-multipose-lightning/resolve/main/onnx/model.onnx
- **File size**: 19,049,169 bytes
- **SHA-256**: `60f724a56f0119d22d2afcf530019db307c77d4ea598eb5bd53bb65afe93930d`

Replaces the site's earlier MoveNet SinglePose Lightning model so the pose estimation demo can
correctly show a skeleton per person on multi-person example images/webcam frames, instead of
only tracking one person and ignoring the rest.

Preprocessing/postprocessing logic used in this site's `pose-estimation.js` was independently
implemented in JavaScript, verified against real image inference (letterbox-pad to 256×256,
int32 RGB pixel input, per-instance keypoint decode + skeleton rendering) and is original code,
not redistributed from the upstream repository.