Today for AI
HOT RADAR
llmHEAT 10.5°

AI CLUSTERED EVENT · 10/14/2026

Shanghai AI Lab Open-Sources InfiniHand: Streaming World-Coordinate Hand Reconstruction at Double the Speed of HaWoR

1 reports archived1 independent sourcesupdated 10/14/2026, 15:47:00
Synthesis & Latest Updates
1 Sources Cross-Validated

Shanghai AI Lab and partners introduced InfiniHand, a unified streaming model that jointly estimates camera trajectories and MANO-parameterized hand poses to map egocentric video actions into a shared world coordinate system. By leveraging a pre-trained LingBot-Map backbone for global geometric context, it eliminates the error accumulation inherent in traditional multi-stage pipelines (detection + SLAM + alignment). Trained on ~5,000 hours of curated public data, InfiniHand achieves 11.19 FPS—more than double the speed of HaWoR—while delivering superior reconstruction accuracy and cross-frame consistency across multiple benchmarks.

LATEST/Shanghai AI Lab and partners introduced InfiniHand, a unified streaming model that jointly estimates camera trajectories and MANO-parameterized hand poses to map egocentric video actions into a shared world coordinate system. By leveraging a pre-trained LingBot-Map backbone for global geometric context, it eliminates the error accumulation inherent in traditional multi-stage pipelines (detection + SLAM + alignment). Trained on ~5,000 hours of curated public data, InfiniHand achieves 11.19 FPS—more than double the speed of HaWoR—while delivering superior reconstruction accuracy and cross-frame consistency across multiple benchmarks.

HEAT TRENDHourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

TIMELINECoverage timeline

Total 1 reports · Latest first
  1. 量子位 (微信公众号)T2·78 pts
    • Architectural Innovation: Uses a single model to jointly estimate camera trajectories and hand poses, replacing traditional multi-system cascades to reduce error propagation and coordination costs.
    • Performance Breakthrough: Achieves an inference speed of 11.19 FPS, more than double that of the leading method HaWoR, meeting real-time requirements for large-scale video processing.
    • Data Engineering: Integrates ~5,000 hours of video from 9 datasets (e.g., ARCTIC, HOT3D), employing unified coordinate definitions and cleaning processes to resolve annotation inconsistencies.
Shanghai AI Lab Open-Sources InfiniHand: Streaming World-Coordinate Hand Reconstruction at Double the Speed of HaWoR | Today for AI