Datasets:

diffracting
/

egocentric-maintenance-sample

Each episode is an excerpt with preserved displayed-source frame identity. Source MOV masters remain locally preserved. Public videos are H.264, 8-bit SDR derivatives of the original HLG HDR capture; resolution is preserved after applying display rotation. Audio and source container metadata are omitted. The documented transform uses Hable tone mapping with npl=100. No image crop or frame-rate resampling is applied to training videos. Browser previews are resized derivatives.

frame_mapping: rgb_frame_index indexes the public training video from zero; source_frame_index indexes displayed frames in the original MOV from zero. This RGB-only source index is not a Stray depth/odometry record index. t_s is the actual public video timestamp relative to its first displayed frame. source_timestamp_s is relative to the original MOV's first displayed frame. source_pts_s and video_pts_s retain native presentation timestamps. Selection is an inclusive start/exclusive end window.
hand_pose_raw, hand_pose: estimated normalized 2D x/y keypoints, track ID, anatomical label or unknown, uncalibrated confidence and provenance. z/wx/wy/wz are null for the 2D model. Temporal interpolation is identified, not presented as an independent detection. Duplicate suppression caps raw detections at two; temporal gap filling can retain short extra tracks.
tracks_raw, tracks: estimated normalized object boxes with labels and confidence. Identity and contact are not ground truth. No physical contact labels are inferred by this recipe.
quality: image statistics, channel coverage and missingness; not annotation accuracy.
annotations/action_segments.parquet, when present: reviewed language/action intervals with explicit review method and approximate boundary precision. Intervals are observations, not robot commands or measured outcomes.

The reader returns None for depth_m, depth_valid, confidence, K_rgb, camera_pose_xyzw and sensor_timestamp_s. Explicit episode modality flags mark metric depth, calibration, VIO, IMU, audio and robot actions unavailable. camera_points raises for RGB-only samples. Estimated monocular geometry is not silently inserted.

All excerpts in the sample are grouped conservatively as one session. Preserve this grouping in downstream splits. Instructions and segment boundaries are curator annotations; no independent accuracy study or task-success benchmark is provided.