---
language: [en, ja]
license: apache-2.0
pipeline_tag: keypoint-detection
---
# Wholebody Video Tracking models
Verified mirrors of Google MediaPipe GHUM Full/Lite Pose and Hand task bundles, plus OpenMMLab RTMPose-s/m 133-point ONNX models.
All original pretrained weights are retained. This project does not retrain on a small image subset.
The default body model is GHUM Full. Lite trades accuracy for compute. RTMPose provides alternate 2D wholebody observations; it does not emit measured depth.
`manifest.json` records SHA-256, original source URLs, model contracts, and immutable Hugging Face asset revisions.
`verification.json` records actual CPU inference on upstream public fixtures. Passing these checks proves executable model contracts, not accuracy on arbitrary videos or low GPU utilization on a particular device.
No user videos, frames, audio, personal landmarks, or credentials are published. Monocular world coordinates and camera-root translation remain estimates.
MediaPipe source: https://github.com/google-ai-edge/mediapipe/tree/v0.10.21 (Apache-2.0).
MMPose source: https://github.com/open-mmlab/mmpose/tree/759b39c13fea6ba094afc1fa932f51dc1b11cbf9 (Apache-2.0).