DewHand is a unified model that directly reconstructs metric world‑space hand motion and camera motion from egocentric (first‑person) video. It predicts both hand and camera motion together in a single stage and in a shared world coordinate system. This avoids the error buildup that can happen when hand and camera motion are estimated separately.

The model resolves three core problems. To fix scale ambiguity from monocular input, DewHand combines scene‑level geometric features from VGGT-Ω with metric‑space supervision. To keep fine hand detail, it preserves hand‑specific features. To handle occlusions and hands moving out of view, it frames reconstruction as conditional motion generation and uses flow matching.

DewHand’s generative formulation outperforms direct regression in both accuracy and temporal smoothness. It achieves these gains with only two sampling steps. In tests on the HOT3D and ARCTIC datasets, DewHand improved world‑space accuracy and temporal consistency versus prior methods. It reduced rigid‑aligned world‑space error relative to StableHand from 32.9 to 28.4 mm on HOT3D and from 76.2 to 36.2 mm on ARCTIC. It also recovers camera scale within 10% on two thirds of HOT3D clips, compared with about a quarter of clips for SLAM.

In zero‑shot evaluations on datasets captured with different devices (HoloAssist, H2O, and TACO), DewHand achieved the lowest rigid‑aligned world‑space error (WA‑MPJPE) on each dataset.

These results support better measurements of first‑person hand motion. That can aid understanding human interactions and transferring motions to embodied agents, as noted by the authors.

Caveats: recovering metric world‑space hand motion from egocentric video remains challenging. The task is complicated by (1) multi‑stage pipelines that propagate errors when hand and camera motion are estimated independently; (2) monocular observations that leave physical scale ambiguous; and (3) occlusion and out‑of‑view hands that allow multiple plausible motion trajectories. Reported improvements are tied to the datasets and comparisons described above.