Noitom Robotics opens HiPHI-Hand hand motion capture dataset, already used to train NVIDIA EgoHand

Embodied data company Noitom Robotics announced the public release on Hugging Face of its high-precision hand motion capture dataset HiPHI-Hand, which has been used to train NVIDIA's 3D hand reconstruction model EgoHand. According to Noitom, EgoHand is built on the NVIDIA Isaac Capture framework and requires only first-person-view video captured by a monocular camera to reconstruct 3D hand pose and shape. The model is released under the OpenMDW license and is permitted for free commercial use. NVIDIA stated that the model was trained on 14 million images, with material drawn from synthetic data, real-world scene videos, and multi-camera motion capture ground-truth data, achieving current state-of-the-art (SOTA) zero-shot accuracy on public first-person-view benchmarks, with a single-frame inference latency of approximately 9 milliseconds.