Capture catalogue · capture brief
Bimanual Household Tasks
This brief is scoped on demand. What that means, and what every brief includes.
What the brief covers
The brief specifies egocentric capture of cooking in Indian household kitchens: cutting vegetables, pouring liquids, stirring and plating. It is built around bimanual work, where one hand holds or steadies while the other cuts, pours or stirs. One Recording can run through several of these in turn; segmentation cuts it into task-bounded Clips, and each Clip keeps its parent's consent receipt. Lighting differs from one kitchen to the next.
The specification
| Capture device | Ray-Ban Meta |
|---|---|
| Camera | 1920x1080 @ 30 FPS |
| Spatial tracks | Camera-relative wrist position (x, y, z), no physical unit |
| Annotation | Tier 1 · Tier 2 · Tier 3 |
| Availability | Scoped on demand |
What a delivery carries
All three FH-Ego v1 tiers apply. Tier 1 describes each Clip: its task label and scene, whether it is usable, its frame range, its consent receipt and its QA verdict. Tier 2 divides it into action segments, each naming the hand that acts (left, right or both), the action and a one-line description. Tier 3 adds a track for each hand on sampled frames: 21 keypoints, wrist position and velocity, and a confidence score. The camera-relative coordinates in the specification are that wrist position, and they carry no physical unit; FH-Ego v1 sets out how each axis is built. A delivery is written as LeRobot v2.1, as HDF5 in the paired layout Apple's EgoDex uses, or as FH-Ego v1's own MP4 and JSON.
Recording it
A household kitchen is a shared room, and other people pass through it while the cook works. Every face the detector finds is blurred, a family member's as much as the cook's, before anyone here opens the footage. The copy a lab receives has no audio track, so a conversation in the next room does not travel with the Clip. The Capture Partner consents on their own phone, on a screen that comes before the record button, and the receipt is in the ledger before the first byte of the Recording reaches us. The glasses put the camera at the wearer's eye line, so the view turns with the head between the counter, the stove and the sink.
What to know before you commission this
- The fifty distinct kitchens named in this brief are the coverage a commissioned engagement would deliver, across varied ambient lighting, stove types and utensils.
- Ray-Ban Meta records 1920×1080 at 30 frames per second from a single camera. Depth and wrist telemetry sit outside this rig; a brief needing either is specified on a different one.
- Clips on this device run to about three minutes and are exported through Meta's own app, and capture sessions are planned around that.
- The sensor is rolling-shutter. A brief that needs global shutter for fast hand motion is specified on the stereo rig.
Sample
Published sample: none