mamma

Why we included this project

MAMMA is a CVPR 2026 oral from the Max Planck Institute for Intelligent Systems and Carnegie Mellon, and it is one of the more complete reference implementations around for markerless multi-person motion capture. Feed it synchronized multi-view video and it reconstructs full-body SMPL-X meshes for several people at once, no markers or suits needed, through a pipeline that combines per-person segmentation, 2D landmark detection, and multi-view optimization. The repo bundles a four-camera example with a quick-start script and a browser UI, so you can run the whole flow on demo footage before adapting it to your own captures. Training code and setup documentation are included as well, which makes it a solid research baseline or a starting point for a markerless mocap project. One thing to plan around: the code and models are released under a non-commercial research license, so commercial use requires a different arrangement.