blanchon/egoinverse-dataset

EgoInverse training corpus

Egocentric video windows with per-frame bimanual hand, camera, body and gravity labels, in the unified shard format read by EgoInverse. This repository is private; access is restricted to the project.

Files

Path	Contents
corpus.json	training manifest (the corpus --corpus-manifest reads)
heldout_corpus.json, heldout_evaluation.json	66 held-out windows (6 per source) and their evaluation manifest
<source>/part-*/	dataset.json, index.jsonl and tar shards of one part
nohand/train/, observable_auxiliary/train/	absence clips and single labelled frames for presence/visibility
heldout/<source>/	held-out parts
benchmark/	the egoinverse benchmark suites: panel/ (54 held-out 121-frame windows of 9 sources), sequences/ (every frame of ARCTIC subject s05 and of HOT3D participants P0010/P0014/P0015, 1,177 windows), visibility/ (18 clips, 90 reviewed frames)

Only parts listed in a manifest belong to the corpus. Each index row records its members' checksums, which the reader verifies. The code's data/unified.py reads this format and training/recipe.py sets the labels each source supervises.

Sources

Source	Samples	Frames	Size	Used for training
ARCTIC	267	186,671	9.5 GB	yes
H2O	114	68,929	3.2 GB	yes
HOT3D (train)	139	34,950	4.0 GB	yes
TACO	1,908	290,112	48.1 GB	camera only
EgoSuite	34,132	15,886,211	179.2 GB	yes
EgoSuite (body clips)	4,878	2,287,725	71.8 GB	yes
EgoDex	328,461	86,855,806	505.0 GB	camera and presence
HoloAssist	1,162	558,922	6.3 GB	yes
OpenAoE	12,091	5,815,426	67.5 GB	yes
STERA	9,431	3,601,789	183.7 GB	yes
GenHumanEgo	2,823	1,357,863	110.4 GB	camera only
NoHand	492	39,852	0.8 GB	yes
Observable auxiliary	998	998	0.2 GB	yes
OakInk2	431	700,624	34.8 GB	evaluation only

Hand labels are motion capture for ARCTIC, H2O, HOT3D, TACO and OakInk2, and estimated for the other sources; the training recipe restricts each source to the supervision its labels support. Body labels (EgoSuite) are estimated, not motion capture.

Terms

Every source keeps its original license and terms of use; several forbid redistribution. Use this corpus only under the terms of each original dataset and of the MANO model license.