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.
This dataset contains depth maps and camera estimates produced with VGGT-Ω, which is released under the FAIR Noncommercial Research License (included as LICENSE_VGGT_OMEGA.txt), together with VITRA-1M annotations. It is derived from Ego4D, EPIC-KITCHENS and Something-Something v2 videos, which have their own license terms. The dataset is released under CC BY-NC 4.0 and may only be used for non-commercial research.
Log in or Sign Up to review the conditions and access this dataset content.
SpatialDex Data
Human-video pre-training data for SpatialDex: Reconstruction-Augmented Visuomotor Pre-Training of 3D-Grounded Bimanual Dexterity from RGB-Only Human Videos (paper, project page).
Per-frame depth and camera estimates from VGGT-Ω, merged with the VITRA-1M episode annotations (MANO hand poses, language instructions, VITRA cameras), for 1,155,175 human-hand episodes (31.7M frames) from Ego4D, EPIC-KITCHENS and Something-Something v2, stored as UniEnv replay buffers.
| dataset | episodes | frames | source videos |
|---|---|---|---|
| ego4d | 948,561 | 23,392,849 | 2,444 |
| epic | 153,896 | 7,222,228 | 580 |
| ssv2 | 52,718 | 1,125,013 | 39,980 |
RGB is not included (the source videos' licenses do not allow redistribution). Every frame carries its source video id and frame index, so RGB can be decoded from the original videos; see Getting RGB.
Files
README.md
LICENSE_VGGT_OMEGA.txt FAIR Noncommercial Research License (VGGT-Ω)
index.parquet one row per episode, for all parts (see below)
ego4d/ego4d_part_0001.tar ... 194 tars
epic/epic_part_0001.tar ... 60 tars
ssv2/ssv2_part_0001.tar ... 12 tars
Each tar (~1–13 GB, uncompressed tar of already-compressed files) extracts to <dataset>/<part>/:
<dataset>/<part>/
<H>x<W>_<NNN>/ a UniEnv ReplayBuffer; one segment per episode, <= 1000 episodes
episodes.parquet the index rows for this part
skipped.jsonl episodes left out and why
DONE
Episodes are grouped by depth resolution, because a buffer has a fixed per-frame shape. The resolution follows
the source video's aspect ratio (~512² pixels, multiples of 16): EPIC 384×688, Ego4D mostly 448×592, SSv2
several. height/width in index.parquet give each episode's size.
Download and extract only what you need, e.g.:
hf download quantumcookie/spatialdex-data --repo-type dataset \
--include "index.parquet" "ssv2/*" --local-dir vitra_vggt
cd vitra_vggt && for t in ssv2/*.tar; do tar -xf "$t" && rm "$t"; done
Per-frame fields
| key | shape / type | meaning |
|---|---|---|
depth |
(H, W) float32, metres | VGGT-Ω z-depth; stored losslessly as FFV1 16-bit video in millimetres (±0.5 mm); 0 = invalid |
camera/pose_c2w |
(4, 4) | VGGT-Ω camera-to-world, OpenCV axes (x right, y down, z forward); world ≈ first camera |
camera/intrinsics |
(3, 3) | VGGT-Ω intrinsics in depth-map pixels (cx = W/2, cy = H/2) |
vitra_camera/extrinsics_w2c |
(4, 4) | VITRA world-to-camera (VITRA's metric world frame, which differs from VGGT-Ω's) |
vitra_camera/intrinsics |
(3, 3) | VITRA intrinsics in raw-video pixels |
hands/{left,right}/kept |
uint8 | 1 if the VITRA hand reconstruction is valid for this frame |
hands/{left,right}/beta |
(10,) | MANO shape (MANO_RIGHT model), constant over the episode |
hands/{left,right}/global_orient_{worldspace,camspace} |
(3, 3) | wrist rotation in the VITRA world / camera frame |
hands/{left,right}/hand_pose |
(15, 3, 3) | MANO local joint rotations |
hands/{left,right}/transl_worldspace |
(3,) | wrist translation, VITRA world, metres |
hands/{left,right}/joints_{worldspace,camspace} |
(21, 3) | 3D joints, metres |
instruction/{left,right} |
text | VITRA description active at this frame ("" if none) |
source/video_id |
text | raw video name (Ego4D uid, EPIC P01_05, SSv2 id) |
source/frame_index |
int64 | 0-based frame index in the raw video (= VITRA video_decode_frame) |
episode_id |
text | <dataset>/<video>_ep_<n> (VITRA's episode naming without the dataset prefix) |
step |
int32 | index within the episode |
VITRA's hand and camera data are metric. VGGT-Ω depth looks metric (typical hand-scene distances 0.5–0.7 m) but its scale is not guaranteed; align it to VITRA if you need exact metric scale.
index.parquet
episode_id, dataset, part, buffer, segment, start, length, height, width, video_id, frame_start, frame_end, vitra_file, anno_type, avg_speed, total_rotvec_degree, total_transl_dist, text, text_rephrase
text / text_rephrase are VITRA's JSON-encoded {hand: [(description(s), (start, end)), ...]} with end
exclusive.
Reading
Requires unienv (pip install unienv, or from GitHub).
import pandas as pd
from unienv_data import ReplayBuffer
from unienv_interface.backends.numpy import NumpyComputeBackend
root = "vitra_vggt"
idx = pd.read_parquet(f"{root}/index.parquet")
ep = idx[idx.dataset == "ssv2"].iloc[0]
rb = ReplayBuffer.load_from(f"{root}/{ep.dataset}/{ep.part}/{ep.buffer}", backend=NumpyComputeBackend)
assert rb.get_segments()[ep.segment] == (ep.start, ep.start + ep.length)
frames = rb.get_at(slice(ep.start, ep.start + ep.length)) # nested dict of arrays, one row per frame
depth, pose = frames["depth"], frames["camera"]["pose_c2w"]
Getting RGB
Decode frames source/frame_index from the raw video source/video_id, then resize the whole frame
(bicubic, no crop) to the depth map's W×H. That is VGGT-Ω's preprocessing (load_and_preprocess_images,
mode="balanced", image_resolution=512), so pixels line up with depth and camera/intrinsics.
Raw-video layouts follow the VITRA data guide:
| dataset | raw video | notes |
|---|---|---|
| ego4d | Ego4D_root/v2/full_scale/{video_id}.mp4 |
VITRA undistorts Ego4D frames first (intrinsics/ego4d/{video_id}.npy in VITRA-1M); it is not yet verified whether VGGT-Ω was run on undistorted frames |
| epic | EPIC-KITCHENS_root/{video_id[:3]}/videos/{video_id}.MP4 |
use as downloaded |
| ssv2 | 20bn-something-something-v2/{video_id}.webm |
use as downloaded |
How it was made
- VGGT-Ω (512 checkpoint) was run on every VITRA-1M episode clip of Ego4D, EPIC-KITCHENS and SSv2, giving per-frame depth, camera-to-world poses and intrinsics.
- Each episode was joined with its VITRA-1M annotation. VGGT-Ω's frame indices equal VITRA's
video_decode_framefor every converted episode; 3 EPIC episodes whose VGGT-Ω output is shorter than the annotation were left out (listed inskipped.jsonl). - Not included: EgoExo4D; ~280 episodes that only exist in an older per-frame VGGT-Ω output format.
- A round-trip check against the source data matched every field exactly except depth (≤ 0.5 mm, the millimetre quantization).
License and terms
- This dataset: CC BY-NC 4.0, non-commercial research only.
- Depth and cameras are outputs of VGGT-Ω, released under the FAIR Noncommercial Research License
(
LICENSE_VGGT_OMEGA.txt), which restricts outputs to non-commercial research and requires acknowledging VGGT-Ω in publications. - Hand poses, VITRA cameras and text come from VITRA-1M (MIT).
- All data is derived from Ego4D, EPIC-KITCHENS (CC BY-NC 4.0) and Something-Something v2 videos; you must comply with their licenses, and you need them to obtain RGB.
Citation
If you use this dataset, please cite SpatialDex, VGGT-Ω and VITRA:
@misc{he2026spatialdex,
title={Reconstruction-Augmented Visuomotor Pre-Training of 3D-Grounded Bimanual Dexterity from RGB-Only Human Videos},
author={Xingyi He and Zhiwen Qiu and Adhitya Polavaram and Yunhao Cao and Cory Fan and Julius Samwer and Saleh Kalantari and Kuan Fang},
year={2026},
eprint={2610.29716},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2610.29716},
}
@misc{wang2026vggtomega,
title={VGGT-$\Omega$},
author={Jianyuan Wang and Minghao Chen and Shangzhan Zhang and Nikita Karaev and Johannes Schönberger and Patrick Labatut and Piotr Bojanowski and David Novotny and Andrea Vedaldi and Christian Rupprecht},
year={2026},
eprint={2605.15195},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2605.15195},
}
@article{li2025vitra,
title={Scalable Vision-Language-Action Model Pretraining for Robotic Manipulation with Real-Life Human Activity Videos},
author={Qixiu Li and Yu Deng and Yaobo Liang and Lin Luo and Lei Zhou and Chengtang Yao and Lingqi Zeng and Zhiyuan Feng and Huizhi Liang and Sicheng Xu and Yizhong Zhang and Xi Chen and Hao Chen and Lily Sun and Dong Chen and Jiaolong Yang and Baining Guo},
journal={arXiv preprint arXiv:2510.21571},
year={2025}
}
Please also cite Ego4D, EPIC-KITCHENS and Something-Something v2 for the underlying videos.
- Downloads last month
- -