minimal-hand

Why we included this project

This project pulls full 3D hand poses out of an ordinary webcam, no depth sensor or extra hardware needed. A detection network finds hand joint locations in a frame, an inverse-kinematics network recovers rotations, and the whole pipeline runs at well over 100fps on a desktop GPU. The design favors simplicity: two stages, pre-trained models that download directly, and a webcam demo you can start right away. The author never released training code and the MANO hand model has to be requested separately, but the paper and community ports to PyTorch and Unity give you enough to learn the approach and build on it. If you want hand tracking for gesture control or interactive AR/VR demos without buying depth hardware, this is one of the cleanest starting points around.