Observation schema v1.0.0
dataset.json indexes complete recordings in the sample subset. Reader RGB indices are zero-based and refer to displayed video frames. source_frame_index identifies the original Stray sensor record and can differ from rgb_frame_index: these recordings contain a discarded pre-roll packet. Both mappings are retained. Browser previews are resized review copies.
frame_mapping: source_frame_index, native video_pts_s, device sensor_timestamp_s, independently zeroed sensor_t_s, and canonical t_s (video PTS minus the first displayed frame PTS), rgb_decodable and rgb_frame_index. Discarded pre-roll records have null RGB indices and may have negative t_s; they remain in the raw sensor tables but are not training RGB samples. Mapping is by explicit source frame ID, not nearest independently zeroed clock. Clock agreement is not an independent physical synchronization measurement.camera_intrinsics: per-frame fx/fy/cx/cy in native RGB pixels, width/height, calibration_source. Downsample focal/principal-point values by depth/RGB resolution ratio for depth unprojection.arkit_poses: t_s, tx/ty/tz in metres and quaternion qx/qy/qz/qw. Camera-to-world in ARKit gravity-aligned world, Y up. Device visual-inertial estimates, not motion-capture truth. OpenCV camera axes are x right, y down, z forward; flip y/z before applying ARKit camera-to-world.sensors/*.zip: original depth PNGs in millimetres and confidence PNGs (0 low, 1 medium, 2 high); CSV sensor payloads preserved byte-for-byte. Reader returns float32 metres and masks confidence <1 and zero depth. Preserve confidence as ordinal levels, not probabilities.device_imu: raw a_x/a_y/a_z and alpha_x/alpha_y/alpha_z (rad/s per source documentation), device timestamps and mapped video t_s. Acceleration units are native/unverified: published docs say m/s^2 but observed magnitudes suggest g-scale logging. No conversion, gravity removal or device-to-camera rotation is assumed. Out-of-coverage t_s is null with clock_mapping_valid=false. This is a phone sensor, not a wrist sensor.hand_pose_rawandhand_pose: 21 estimated 2D keypoints per hand, normalized x/y, track ID, handedness, scores. z/wx/wy/wz are null for the 2D detector. Temporal outputs carry detected/smoothed/interpolated/extrapolated provenance. Detection scores are not calibrated error probabilities.hand_world_trajectory: depth-anchored estimated wrist position; optional estimated palm-plane xyzw orientation, orient_valid. Orientation null when unavailable. Metric coordinates do not imply measured anatomical accuracy. The 2D detector and low-resolution depth are error sources. The single-wearer recipe suppresses duplicate candidates, caps raw detections at two hands (temporal gap filling can retain additional short tracks), and marks conflicting model handedness unknown; original candidate selection counts are disclosed in quality metadata.tracks*,contact*: experimental object boxes and hand/object proximity estimates. Cutting deforms objects; instance identity and true physical contact are not ground truth. Contact has no force/tactile supervision.quality/*.json: completeness, coverage, missingness and processing metadata. No aggregate accuracy score. No robot action, reward or terminal-success field is fabricated.
Signals retain their native sampling rates. A 60 fps RGB stream does not mean 60 independent hand estimates per second. Use bounded joins (annotations_near) and preserve absent values. Never forward-fill across long gaps.