metrabs

Why we included this project

MeTRAbs is built around a question most pose estimators skip: where a person's joints are in real, metric-scale 3D space, not just where they appear on a flat image. It works from ordinary RGB images, and the output carries actual distances and positions, which is what you need for motion capture, sports analysis, ergonomics, and similar work. The models load as TensorFlow SavedModels through TensorFlow Hub, and the repo includes a straightforward inference API plus Colab notebooks and demo scripts for images, video, and webcam input, so you can move from setup to running predictions quickly. Multiple skeleton conventions and camera distortion handling make it practical to integrate into an existing pipeline. The code comes from peer-reviewed research and offers backbone choices from fast mobile-friendly models to higher-accuracy ones, letting you trade speed against precision.