EgoHandBench

EgoHandBench is a benchmark for egocentric 3D hand reconstruction. It converts seven public datasets into the same format, removes frames with low-quality annotations, and fixes the train/test splits. It also provides the evaluator and ready-to-run baselines.

Code, evaluation and leaderboard are on GitHub.

Datasets

Dataset	Config	Setting	Train clips	Test clips	Test frames	Resolution	Source
ARCTIC	arctic	In-domain	277	37	25,783	672 × 480	arctic.is.tue.mpg.de
H2O	h2o	In-domain	180	54	30,439	832 × 480	h2odataset.ethz.ch
HOT3D	hot3d	In-domain	562	29	25,740	480 × 480	HOT3D
OakInk2	oakink2	In-domain	680	258	237,591	832 × 480	OakInk2
EgoDemo Pro	egodemo_egopro_standard	In-domain	1,191	184	195,170	608 × 480	LightwheelAI/EgoDemo
EgoDemo Standard	egodemo_egostandard	In-domain	727	80	156,250	608 × 480	LightwheelAI/EgoDemo
Gen-HumanEgo	gen_humanego_1h	In-the-wild (test only)	–	46	109,580	576 × 480	genrobot2025/Gen-HumanEgo

The frames excluded from each source clip, with the reason for each, are listed in datasets/_trace on GitHub.

Layout

datasets/{dataset}/
  mp4/rgb_480p/{clip}.mp4      RGB video
  cameras_json/{clip}.json     per-frame camera intrinsics
  mano/{clip}.npz              per-frame MANO ground truth for both hands
  manifests/                   split.json, train.parquet, test.parquet
baselines/{method}/{official,egohandbench}/   weights used by the GitHub baselines

All hands use the same MANO convention (MANO_RIGHT, with left hands stored mirrored). official weights are the authors' releases, and egohandbench weights are retrained on the train splits above.