// the find
rwightman/posenet-pytorch
A PyTorch port of Google TensorFlow.js PoseNet (Real-time Human Pose Estimation)
A straight PyTorch port of Google's TensorFlow.js PoseNet model for multi-pose human keypoint detection. Useful if you want a lightweight, dependency-light pose estimator without pulling in TF.js, but it's explicitly unmaintained and the author redirects people to a community fork.
Converts TF.js PoseNet weights on the fly so you don't need to hand-port weights yourself. Ships three working demos (image, webcam, benchmark) that run out of the box rather than just a model definition. Decode step is partially moved to GPU, giving a real throughput number (130fps on a 1080 Ti) instead of a vague performance claim.
Archived by the author, who points you to a different fork for anything beyond inference — no fixes or updates are coming here. The author's own TODO admits there's no rigorous verification that outputs match the original TensorFlow.js model, which matters a lot for a pose estimation library. Pinned to old, specific versions (PyTorch 1.0, opencv-python 3.4.5.20) because newer opencv builds broke drawKeypoints — expect dependency pain on a modern setup. No training or fine-tuning code, no batch inference, and no tests of any kind.