InfiniHand: Streaming World-Space Hand Motion Estimation from Egocentric Video
InfiniHand jointly estimates MANO parameters, camera trajectories, and hand locations from uncalibrated egocentric video.
InfiniHand is a feed-forward streaming model that jointly estimates MANO hand parameters, camera trajectories, and world-space hand locations from uncalibrated egocentric video, replacing cascaded pose estimators and SLAM. It is trained in two stages on about 5,000 hours of public egocentric data, first learning camera-space hand priors and then streaming world-space reconstruction. On in-domain benchmarks it cuts ARCTIC PA-p by 21.4% versus ViDiHand, reduces world-space drift, and runs at 11.19 FPS, more than twice HaWoR.
- End-to-end streaming model jointly predicts MANO parameters, camera motion, and hand locations.
- Pretrained on roughly 5,000 hours of public egocentric video.
- Reduces ARCTIC PA-p by 21.4% versus ViDiHand and runs at 11.19 FPS.
- More than twice HaWoR throughput while reducing world-space drift.
From the original article
1 min read at huggingface.co
World-space hand motion estimation from egocentric video requires recovering 3D articulated hand geometry while tracking camera egomotion. Existing approaches heavily rely on cascading independent hand pose estimators and SLAM systems, resulting in error accumulation, complex pipelines, and severe computational overhead.
By Kerui Ren, Kaiwen Song, Weiguang Zhao · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.