Reading muscle activity directly from video

Understanding human movement requires more than tracking skeleton position—it demands knowledge of internal muscle activity, essential for rehabilitation and injury prevention. This work introduces BioHuman10M, a large-scale dataset pairing video with motion capture and simulated muscle activations, plus BioHuman, an end-to-end model that predicts both kinematic motion and muscle activations from single-camera video. The model generalizes across subjects and motion types, bridging the gap between visual observation and biomechanical state that prior methods could not directly infer.