Shanghai AI Lab's InfiniHand Tracks 3D Hands in World Space at 11 FPS from Egocentric Video

Shanghai-AI-Laboratory · hf · 2026-09-29

Shanghai AI Lab presents InfiniHand, an end-to-end streaming feed-forward framework that jointly estimates MANO hand parameters, camera trajectories, and world-space hand locations from uncalibrated egocentric video, replacing cascaded pose-estimator+SLAM pipelines. Trained on 5,000 hours of egocentric data in two progressive stages, it cuts ARCTIC PA-p by 21.4% vs ViDiHand, mitigates world-space drift, and runs at 11.19 FPS — over 2x HaWoR's throughput.

Original post →

More from Embodied

Embodied channel →