diffracting/egocentric-kitchen-sample

Diffraction Egocentric Kitchen Capture Sample

A small, inspectable sample of human kitchen manipulation captured with Stray Scanner on a LiDAR-equipped iPhone: native RGB, metric depth and confidence, per-frame camera calibration, device odometry, raw device IMU, and explicitly estimated hand/object annotations.

Human observation sample. License: cc-by-4.0. This sample contains 3 recordings totaling 167.85 seconds. It is an observation dataset for evaluating human-video pretraining, spatial perception, action understanding and motion-prediction workflows. It supplies no measured robot commands, force or tactile data, and demonstrates no robot-policy performance gain.

Capture	Seconds	Native RGB frames	Instruction
onion-preparation	75.73	4,543	Prepare and peel the onion on the cutting board using hands and a kitchen knife.
pepper-slicing	57.75	3,465	Prepare and slice the bell pepper on the cutting board using a kitchen knife.
pepper-dicing	34.37	2,062	Dice the bell pepper strips on the cutting board using a kitchen knife.

Download and try one episode

pip install huggingface_hub
hf download diffracting/egocentric-kitchen-sample --repo-type dataset --local-dir diffraction-sample
cd diffraction-sample
pip install -r requirements.txt
python observation_reader.py .

The download includes native RGB-D recordings and a portable reader. The Hugging Face dataset viewer uses smaller video previews; use videos/ for native RGB training pixels.

from observation_reader import ObservationDataset, camera_points
sample = ObservationDataset('.')
key = next(iter(sample.episodes))
frame = sample.frame(key, 100)
print(frame['rgb'].shape, frame['depth_m'].shape, frame['K_rgb'])
points = camera_points(frame)
hands = sample.annotations_near(key, 'hand_pose', frame['timestamp_s'], max_age_s=0.05)

Native videos are under videos/; smaller browser previews and episode metadata under previews/. Depth/confidence and original sensor CSVs are in sensors/. Annotation tables are standalone Parquet. SCHEMA.md defines units, frames, clock mapping, confidence and missing-data behavior. checksums.sha256 covers every packaged artifact.

Collection and scope

The owner confirmed these Stray Scanner recordings are owned and may be used to prepare a training-data sample. This small collection shows closely related kitchen tasks; participant count, device model/app version and independent session identity were not supplied. All recordings belong to a conservative common session group. There is no claimed train/test benchmark or cross-person generalization.

The pipeline preserves native video bytes and joins camera/depth records by source frame ID, retaining both original clocks. Per-frame intrinsics are used for wrist lifting. Raw and temporally processed keypoints are separate. Instructions were prepared from video inspection; detailed action boundaries, success/failure and ground-truth contact annotations are not provided.

Quality and limitations

Per-episode quality/ reports expose channel coverage, null counts, provenance and sensor correspondence. These are engineering checks, not labeled accuracy benchmarks. ARKit camera pose is a device VIO estimate. Hand pose, palm orientation, object tracking and contact remain estimates; review their overlays and confidence/missingness before training. The low-resolution depth sensor can miss thin fingers, tool edges and occlusion boundaries. Fast cutting, bimanual occlusion and deformable vegetables are difficult cases.

Raw IMU acceleration units need capture-version verification; no silent unit conversion is applied. Device IMU is not wrist-motion sensing. See SCHEMA.md. An independent labeled evaluation and robot transfer experiment are future work, not measured results of this sample.

Reproducibility and professional capture

The accompanying processing record identifies the source code and environment used. QUALITY.md summarizes observed coverage and limitations; validation.json records the portable-reader and file-integrity checks. Run python validate_release.py . after installing datasets alongside the reader requirements to repeat these checks. This sample is intended to start technical evaluation of task-specific professional capture: agree on target activities, environments, collection protocol, annotation requirements and measurable acceptance criteria before scaling a collection.

References

Stray Scanner source format, Hugging Face video datasets.

Attribution

Please attribute Diffraction, Egocentric Kitchen Capture Sample (2026) and include the dataset URL and the revision you used. The dataset is licensed under CC BY 4.0.