OmniEgoCap: Camera-Agnostic Sequence-Level Egocentric Motion Reconstruction
Paper • 2512.19283 • Published
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
OmniEgoDB is released under CC BY-NC 4.0 for non-commercial research. The recordings show the four consenting participants and, occasionally, bystanders whose faces have been blurred. By requesting access you agree to use the data for non-commercial research only, not to attempt to identify or re-identify any person in the data, and to delete data on request if a person in it asks for removal.
Log in or Sign Up to review the conditions and access this dataset content.
OmniEgoDB
🌐 Project Page | 📜 Paper | 💻 Code
OmniEgoDB is a multi-device egocentric motion capture benchmark: Xsens ground-truth full-body motion paired with Project Aria, GoPro and Meta Quest recordings. 50 sequences, 4 subjects, 177.2 minutes.
| Device | Sequences | Minutes | Frames (30 fps) | Size |
|---|---|---|---|---|
| Aria | 21 | 67.1 | 120,733 | 75.9 GB |
| GoPro | 20 | 68.1 | 122,673 | 25.7 GB |
| Quest | 9 | 42.0 | 75,638 | 1.4 GB |
Layout
OmniEgoDB/
├── aria/<sequence>/
│ ├── input.npz model input, wrists from Aria MPS hand tracking
│ ├── input_rgb.npz model input, wrists estimated from the RGB video (HaWoR)
│ ├── gt.npz ground truth on the same frames
│ ├── gt.mvnx Xsens MVN export (60 Hz)
│ ├── recording.vrs Aria recording (blurred)
│ └── mps/ Project Aria Machine Perception Services output (SLAM, hand tracking)
├── gopro/<sequence>/
│ ├── input.npz model input, wrists estimated from the video (HaWoR)
│ ├── gt.npz
│ ├── gt.mvnx
│ ├── video.MP4 GoPro video (blurred)
│ └── camera.json intrinsics
└── quest/<sequence>/
├── input.npz model input, head and wrists from the headset's tracking
├── gt.npz
├── gt.mvnx
└── poselog.csv raw head and hand tracking log
Sequence names are S0<subject>_<location>_<index>, e.g. S03_office_2.
Files
World frame: meters, z up, floor at z = 0, 30 fps. Poses are 4×4 local-to-world transforms.
input.npz / input_rgb.npz — input to OmniEgoCap
| Key | Shape | |
|---|---|---|
head |
(T, 4, 4) | head (SMPL-H head joint) |
wrist |
(T, 2, 4, 4) | left / right wrist |
visible |
(T, 2) | whether each wrist is observed |
device_pose |
(T, 4, 4) | device pose |
points |
(N, 3) | scene point cloud (Aria, GoPro) |
raw_time_s |
(T,) | time on the raw recording |
gt.npz — ground truth
| Key | Shape | |
|---|---|---|
segments |
(T, 23, 4, 4) | Xsens segment poses, same world and frames as the input |
segment_names |
(23,) | segment order |
valid |
(T,) | frames to score |
meta |
() | JSON; frame t is mvnx frame start_xsens + 2t |
camera.json (GoPro) is an OpenCV model of the 3840×2160 video.
Evaluation
Use the OmniEgoCap code. --track is one of aria, aria_rgb,
gopro and quest.
python scripts/eval.py omniegodb --root /path/to/OmniEgoDB --track aria
python scripts/visualize.py --omniegodb /path/to/OmniEgoDB --track aria
Privacy
All four participants consented to the release. Faces of other people and license plates are blurred. Aria eye tracking and audio, GoPro audio and GPS, and participant names were removed. The ground truth and inputs were computed from the unblurred originals, so we recommend masking blurred regions when running SLAM on the videos.
License
CC BY-NC 4.0. RGB-based wrists (aria/*/input_rgb.npz,
gopro/*/input.npz) were estimated with HaWoR (CC BY-NC-ND 4.0, MANO).
SMPL-H and MANO are not included. Do not attempt to identify any person in the data. For removal requests, contact
cscandkswon@snu.ac.kr.
Citation
@inproceedings{cho2026omniegocap,
title = {OmniEgoCap: Device-Agnostic Sequence-Level Egocentric Motion Reconstruction},
author = {Cho, Kyungwon and Na, Jeonghyeon and Joo, Hanbyul},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
year = {2026}
}
If you use the Aria recordings, please also cite Project Aria:
@misc{engel2023projectarianewtool,
title = {Project Aria: A New Tool for Egocentric Multi-Modal AI Research},
author = {Jakob Engel and Kiran Somasundaram and Michael Goesele and Albert Sun and Alexander Gamino and Andrew Turner and Arjang Talattof and Arnie Yuan and Bilal Souti and Brighid Meredith and Cheng Peng and Chris Sweeney and Cole Wilson and Dan Barnes and Daniel DeTone and David Caruso and Derek Valleroy and Dinesh Ginjupalli and Duncan Frost and Edward Miller and Elias Mueggler and Evgeniy Oleinik and Fan Zhang and Guruprasad Somasundaram and Gustavo Solaira and Harry Lanaras and Henry Howard-Jenkins and Huixuan Tang and Hyo Jin Kim and Jaime Rivera and Ji Luo and Jing Dong and Julian Straub and Kevin Bailey and Kevin Eckenhoff and Lingni Ma and Luis Pesqueira and Mark Schwesinger and Maurizio Monge and Nan Yang and Nick Charron and Nikhil Raina and Omkar Parkhi and Peter Borschowa and Pierre Moulon and Prince Gupta and Raul Mur-Artal and Robbie Pennington and Sachin Kulkarni and Sagar Miglani and Santosh Gondi and Saransh Solanki and Sean Diener and Shangyi Cheng and Simon Green and Steve Saarinen and Suvam Patra and Tassos Mourikis and Thomas Whelan and Tripti Singh and Vasileios Balntas and Vijay Baiyya and Wilson Dreewes and Xiaqing Pan and Yang Lou and Yipu Zhao and Yusuf Mansour and Yuyang Zou and Zhaoyang Lv and Zijian Wang and Mingfei Yan and Carl Ren and Renzo De Nardi and Richard Newcombe},
year = {2023},
eprint = {2308.13561},
archivePrefix = {arXiv},
primaryClass = {cs.HC}
}
- Downloads last month
- 99