Sample quality and validation
This release contains three owned kitchen recordings, totaling 167.85 seconds and 10,070 displayed RGB frames at 1920 x 1440. Native depth/confidence images are 256 x 192. All recordings remain in one conservative session group; they are a technical sample, not a train/test benchmark.
Sensor integrity
Every displayed RGB frame has a corresponding source frame ID, depth/confidence pair, camera pose and per-frame intrinsic calibration. Each source also has one discarded pre-roll video packet; its original sensor record is retained with a null RGB index. This explains the difference below. The original clocks, video bytes and sensor payloads are preserved. Camera pose is a device VIO estimate, not independent ground truth. Raw IMU acceleration units remain unverified for the capture app version.
| Recording | Displayed RGB | Sensor records | Unknown hand label rows | Timestamps with >2 temporal hands |
|---|---|---|---|---|
| onion-preparation | 4,543 | 4,544 | 69.7% | 49/1136 |
| pepper-slicing | 3,465 | 3,466 | 95.6% | 26/866 |
| pepper-dicing | 2,062 | 2,063 | 95.2% | 0/515 |
Optional annotation quality
The RGB-D capture is the primary data. RTMPose/RTMW hand estimates, OWLv2 object boxes, temporal tracks, depth-lifted wrists and contact estimates are experimental auxiliary channels. Raw detections are capped at two hands after duplicate suppression; temporal gap filling can retain short extra tracks. The counts above expose this behavior. Do not treat track IDs as verified anatomical identity. Conflicting left/right predictions are unknown.
Sampled visual inspection of three evenly spaced points in every final overlay found useful detections alongside occlusion errors, loose object boxes and ambiguous hand identity. No independent keypoint, contact or tracking accuracy benchmark has been conducted. Per-frame scores are not calibrated error probabilities. Missing 3D hand outputs remain null; palm orientation is frequently unavailable. Annotated overlays are provided under review/ for inspection and do not alter native training pixels.
Validation evidence
validation.json records actual consumer checks: a standalone reader in an isolated environment decoded five positions in every video, including the first and last; it loaded the corresponding depth/confidence/calibration/pose, checked geometry and missing-value semantics, loaded the three video rows with Hugging Face Datasets, and verified packaged SHA-256 hashes. These checks establish packaging integrity, not annotation accuracy or robot transfer. source_preservation.json compares released video and sensor payloads with the three approved source archives.
An automated full-video face scan returned zero face detections for all three recordings. Sampled visual inspection found kitchen work surfaces and the wearer's hands. This is not an independent human privacy audit. The owner authorized these recordings and selected CC BY 4.0.
Next evidence to collect with a lab
A pilot should specify activities, participant/session diversity, device and mount metadata, environment conditions and the acceptance criteria before capture. A small independently labeled subset can quantify hand visibility, keypoint error, object-track stability and action boundaries. A downstream ablation can then test whether native RGB-D and optional annotations improve the lab's target task. These results are not claimed by this sample.