JulienXAstrid/egoinverse-dataset

EgoInverse training corpus

This is the data EgoInverse was trained and benchmarked on: about 118 million frames of video filmed from people's heads, from 14 sources, with labels for both hands, the camera, gravity and, for some clips, the body. The weights are at JulienXAstrid/egoinverse.

hf download JulienXAstrid/egoinverse-dataset --repo-type dataset --local-dir data

What's inside

Source	Samples	Frames	Size	Hand labels
ARCTIC	267	186,671	9.5 GB	motion capture, MANO
H2O	114	68,929	3.2 GB	motion capture, MANO
HOT3D (training people)	139	34,950	4.0 GB	motion capture, MANO
TACO	1,908	290,112	48.1 GB	motion capture, MANO
OakInk2	431	700,624	34.8 GB	motion capture, MANO
EgoSuite	34,132	15,886,211	179.2 GB	estimated joints
EgoSuite (body clips)	4,878	2,287,725	71.8 GB	estimated joints and a 22-joint body
EgoDex	328,461	86,855,806	505.0 GB	headset hand tracking
HoloAssist	1,162	558,922	6.3 GB	headset hand tracking
OpenAoE	12,091	5,815,426	67.5 GB	model predictions
STERA	9,431	3,601,789	183.7 GB	estimated
GenHumanEgo	2,823	1,357,863	110.4 GB	pseudo-labels
NoHand	492	39,852	0.8 GB	none: clips without hands
Observable auxiliary	998	998	0.2 GB	single frames marked "hand visible" or not

Camera poses come from each dataset's tracking (headset, rig or capture system). Every clip also carries a gravity direction where it can be recovered.

How each source is used in training

No source has everything. Motion capture is exact but covers few people and places; the large egocentric datasets cover every kind of scene but their hands are estimates. So each source only trains what its labels are good for.

• Motion capture (ARCTIC, H2O, HOT3D) trains everything, including hands that leave the frame. Some training views are deliberately cropped away from the hands.
• Estimated hands (EgoSuite, OpenAoE, STERA) don't have a reliable hand size, so the prediction is rescaled to the label before positions are compared. They teach pose and appearance without teaching a wrong scale.
• Headset tracking (EgoDex, HoloAssist) teaches camera motion and when hands are visible. EgoDex's hand geometry and intrinsics are not used.
• Camera only (TACO, GenHumanEgo): their hands are not used.
• NoHand and the observable frames teach the model when not to draw a hand.
• OakInk2 is kept for evaluation.

The exact recipe is SOURCES in training/recipe.py, and the losses behind each column are described in docs/architecture.md.

Files

Path	What it is
corpus.json	the training manifest: which parts make up each source, and which labels they must have
<source>/part-*/	one part: dataset.json, index.jsonl and the tar shards
heldout_corpus.json, heldout_evaluation.json, heldout/	66 held-out clips scored during training
benchmark/	the clips used by egoinverse benchmark

Format

Each sample in a tar shard is three files:

• <key>.video.mp4: the clip, H.264 at near-lossless quality (CRF 10), at most 832 pixels wide, at the source frame rate.
• <key>.labels.npz: the labels, frame by frame. Hand joints are in metres in the camera frame (x right, y down, z forward), 21 joints per hand in OpenPose order, left hand first, with a validity mask. Alongside: the intrinsics, camera poses, MANO parameters, hand visibility, gravity and, where available, the body.
• <key>.meta.json: where the clip comes from (recording, person, frame range, frame rate).

index.jsonl lists every sample with the byte offset and size of each file. dataset.json states the conventions.

Loading with the code

from pathlib import Path
from egoinverse.data.corpus import Corpus

corpus = Corpus(Path("data/corpus.json"))
arctic = corpus.source("arctic")
episode = arctic.episodes()[0]
clip = arctic.load(episode, start=0, num_frames=81)   # frames, hands, MANO, camera, ...

File	Role
data/corpus.py	reads manifests and sources
data/unified.py	reads shards and checks them
data/record.py	the ClipRecord a clip loads into
training/data.py	samples and augments training clips

Benchmark

egoinverse benchmark --checkpoint model/model.safetensors --data data/benchmark --output results

Suite	What it is
benchmark/panel/	54 held-out clips of 121 frames from 9 datasets
benchmark/sequences/	every frame of ARCTIC subject s05 (25,203) and of HOT3D participants P0010, P0014 and P0015 (71,400)
benchmark/visibility/	16 short videos with 80 reviewed frames: gloved, bare, hidden and hand-free hands

Terms

Each source keeps its original license and terms of use, and several forbid redistribution. Use this data only under the terms of each original dataset and of the MANO license.