FactorizedHMR: A Hybrid Framework for Video Human Mesh Recovery

In three linesFactorizedHMR introduces a two-stage hybrid framework for video human mesh recovery. A deterministic regression module anchors torso and root, while a probabilistic flow-matching module completes ambiguous distal articulations (arms, legs). Geometry-aware supervision and classifier-free guidance improve recovery under occlusion.

Summary generated by Claude — human-verified