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.
EgoPeekaboo is derived from Xperience-10M (huggingface.co/datasets/ropedia-ai/xperience-10m), a controlled-access dataset with its own agreement. The RGB video, camera poses, IMU and hand motion capture in EgoPeekaboo come from Xperience-10M. Request access only if you have accepted the Xperience-10M terms; your use of EgoPeekaboo is subject to them.
Log in or Sign Up to review the conditions and access this dataset content.
EgoPeekaboo
EgoPeekaboo is egocentric 4D ground truth for point tracking. It covers 3,740 Xperience-10M episodes: 253 hours at 10 fps. For each frame it has the RGB image, metric depth and the camera pose, plus 3D tracks of moving objects and people and MANO meshes of both hands. It is stored as contiguous 96-frame chunks. A reader cuts training windows out of them and draws query points when it loads, as in D4RT.
One rule runs through the data: what is not accurate is not used. Every label has a validity mask. Where a depth value, a track position or a hand vertex cannot be trusted, its mask is 0 and nothing is filled in.
Contents
| Episodes | Chunks | Frames | Hours | Size | |
|---|---|---|---|---|---|
| phase1 | 2,510 | 56,362 | 5,529,982 | 153.6 | 391.7 GB |
| expand100 | 1,230 | 36,700 | 3,582,621 | 99.5 | 251.7 GB |
| all | 3,740 | 93,062 | 9,112,603 | 253.1 | 643.4 GB |
22 phase1 episodes are shorter than 96 frames. Each of them is one file under short_episodes/. 7 expand100
episodes are left out because their depth could not be calibrated; they are listed under excluded in the manifest,
with the reason.
README.md
docs/FORMAT.md every field, code and convention
loader/ the reader (Python); loader/requirements.txt
examples/read_pair.py read two adjacent chunks as one 192-frame sample
validation/ upload and data checks
xperience/v4.2/
manifest.json every episode: ordered chunks with frame range, size and SHA256
manifest.sha256
chunks/<capture>_ep<n>/<start>-<stop>-<hash>.h5
short_episodes/chunks/<capture>_ep<n>/...
How the labels are made
- Depth. The Xperience stereo depth of each frame is first calibrated with one correction per episode, fitted
to the SLAM landmarks. The calibrated frames are then fused into one TSDF surface per episode.
- A static pixel takes the depth of that surface rendered into the frame, and only where the frame's calibrated stereo agrees with it within max(2 cm, 2 %). Elsewhere the pixel has no depth (0). Because every frame renders the same surface, a static point keeps the same 3D position across the whole episode.
- Moving objects, people and hands keep the calibrated stereo depth. An object's depth is 0 in a frame where at least half of its tracks jump at once, and throughout a window where its 3D is not trusted (see the next item).
- Moving objects and people. GMOS motion segmentation gives the instances, and AllTracker tracks points on them.
- A track's 3D is the calibrated depth at its pixel, smoothed along the camera ray. The jumps in the stereo depth lie along the ray, so only the distance along the ray is re-estimated.
- A frame's 3D is left out (
xyz_valid0) when the depth jump flags fire, when the point is more than 3 m from the camera, or when the smoother marks it as an outlier. - If an object's depth jumps within a 10 s window, all of its tracks lose their 3D in that window.
- Where a corrected re-identification shows that a track has moved onto another object, the track stops supervising.
- Hands. The hands are MANO meshes from the Xperience hand motion capture: 778 vertices per hand, in world coordinates. A vertex supervises a frame where it is visible and the depth confirms it within max(3 cm, 5 %). It counts as hidden where the depth lies clearly in front of it.
- Static background. Any pixel outside the motion masks that has static depth is a static candidate. The reader
lifts it to 3D with the depth and the pose, and projects it into the other frames, where the depth decides whether
it is visible. The reader leaves a static point out when:
- it comes within 15 cm of a hand in the window (what the hands touch may move);
- its target camera is more than 30 cm from its source camera;
- an object mask covered it within 2 frames or 2 px (masks flicker).
Checks
The numbers are in validation/.
- Static ground truth against the images. We matched the source and target frames of held-out static queries with RoMa v2 and, independently, with an NCC patch match. RoMa lands a median 0.1β0.7 px from the ground truth. In training-like windows, both matchers miss the ground truth and agree with each other (a confirmed error) for 0.32 % of the static pairs.
- Dynamic smoothness. This was measured over every episode, on supervised track points with valid 3D in three consecutive frames. It is the out-and-back step along the ray, |Οt β (Οtβ1 + Οt+1)/2|. Its share at or above 2 / 5 / 10 cm is 0.35 / 0.077 / 0.019 % for phase1 and 0.28 / 0.072 / 0.017 % for expand100.
- Integrity. Every chunk's size and SHA256 match the manifest, both on disk and on the Hub. The reader was run
from
loader/alone, outside our code base.
Reading
pip install -r loader/requirements.txt
import sys
sys.path.insert(0, "loader")
from vlm_track.annotate.d4rt_dataset import XperienceChunkDataset, XperienceD4RTConfig
ds = XperienceChunkDataset("xperience/v4.2/manifest.json", # a downloaded copy, or hf://datasets/2inf/EgoPeekaboo/...
cache_dir="/local/ssd/egopeekaboo_cache", config=XperienceD4RTConfig())
sample = ds[0]
The reader takes a local path or an fsspec URI, including hf:// with an approved account. It downloads a chunk
only when a sample needs it, checks the chunk's SHA256, and keeps it in a bounded node-local cache. To download
everything, run hf download 2inf/EgoPeekaboo --repo-type dataset --local-dir <dir>.
XperienceD4RTConfig() sets the defaults:
- 48-frame windows at stride 1β3, 256 px.
- 2,048 queries: 50 % static, 35 % object and 15 % hand. When a kind has too few candidates, its share goes to the others. Queries are drawn without repetition while the window has enough distinct candidates.
- 30 % of the queries come from depth boundaries.
A sample holds the frames, poses and intrinsics; the queries (source pixel and frame, target frame); and
their targets (2D position, 3D position, visibility). Each target has its own validity mask. The sample is in a
gravity-aligned frame, with world up [0, -1, 0]. docs/FORMAT.md describes every field.
Terms
EgoPeekaboo is derived from Xperience-10M. Its RGB video, camera poses, IMU and hand motion capture come from there. Access is gated, and use is subject to the Xperience-10M terms. No other license is granted.
- Downloads last month
- 51