convolutional-pose-machines-tensorflow

Why we included this project

Convolutional Pose Machines is a TensorFlow implementation of the staged-heatmap approach to 2D body and hand keypoint estimation, kept runnable on its own rather than buried inside a heavyweight framework. It comes with demo scripts for both body and hand pose, accepts live webcam input as well as image and video files, and adds a few practical extras like Kalman-filter smoothing and a simple single-hand tracker. The multi-stage graph construction is easy to follow, which makes it a reasonable place to start if you want to study how staged heatmap refinement works or retrain on your own joint-annotation data. Two caveats: it targets TensorFlow 1.4, and the model assumes the subject is roughly centered in the frame, so real-world scenes usually mean pairing it with a detector up front. There's also a distillation training scheme for compressing a larger teacher network into a smaller student model.