Sample package
What a delivered sample looks like.
An illustration of the structure and records that ship with every dataset. The numbers and clips below are examples, not a released dataset.
Illustrative
Egocentric manipulation, sample package.
Example layout for a small first-person manipulation sample.
Example sample packageNot a released dataset
- Task
- Tabletop pick, place and pour, performed by a human wearing a head-mounted camera
- Streams
- RGB, depth, hand pose, IMU
- Episodes
- Scoped with you; a sample is typically a small set per task variant
- Labels
- Sub-task boundaries, object instances, success or failure, with failures kept
- Splits
- Train, validation and held-out evaluation from separate sessions and operators
- Delivery
- MP4, Zarr and Parquet, JSON annotations, calibration files, QA report
Contents
Every package ships with its own record.
| Path | What it holds |
|---|---|
episodes/ep_0001/rgb.mp4 | Head-mounted RGB, 1080p, 30 fps |
episodes/ep_0001/depth.zarr | Depth aligned to RGB, 16-bit, metres |
episodes/ep_0001/hands.parquet | Per-frame hand pose and contact flags |
episodes/ep_0001/imu.parquet | Head IMU, timestamped on the capture clock |
episodes/ep_0001/annotations.json | Sub-task boundaries, object instances, outcome |
calibration/ | Intrinsics, depth-to-RGB extrinsics, per-session check |
provenance.json | Device, operator ID, consent record, reviewer, guideline version |
qa_report.pdf | Review pass rate, rejection reasons, inter-reviewer agreement |
Provenance
An example provenance record.
provenance.json (example values)Illustrative
- device
- Head-mounted camera, serial and firmware logged
- calibration
- Checked at the start of each session; drift flagged
- operator
- Anonymous ID, with a signed consent and release on file
- reviewer
- Two-pass review; rejected clips recorded with the reason
- guideline
- Version of the labeling guideline used for the clip