pengrui-rio/FEAGINE_HumanEgo_DataSample

EgoData Sample 1520P

Current collection

59 episodes · 139,836 frames · 77.687 minutes · 322 timed subtasks.

All counts and statistics on this page cover the entire v5 collection, sample_001 through sample_059. They were recomputed from the delivered metadata and Parquet files on 14 October 2026; no episode subset is used.

Each episode provides six first-person RGB views in three stereo pairs, at 1920 × 1520 and 30 FPS, encoded as H.264. Images are undistorted pinhole views with corresponding camera intrinsics. The data includes 21 joints per hand, wrist and MANO parameters, world-coordinate trajectories, camera poses, and timed English task annotations.

Valid 3D trajectory coverage

Measurement	All-frame coverage	Median episode coverage	Minimum episode coverage
Left hand	98.00%	99.06%	90.59%
Right hand	98.34%	99.14%	92.97%
Both hands simultaneously	96.65%	97.30%	90.18%

A valid hand/frame has both frame_valid and trajectory_valid set, all 21 joint_valid flags set, and finite joint coordinates. Missing rows count as unavailable. All-frame coverage uses all 139,836 frames as its denominator; the episode minimum and median weight each episode equally. This measures valid trajectory availability, not actual image visibility or recognition accuracy.

Frame-to-frame motion

Measurement	Mean	Median	P95
Wrist translation (mm/frame)	3.76	2.20	12.23
Wrist orientation (degrees/frame)	2.08	1.37	6.44
Reference-camera translation (mm/frame)	1.29	0.60	4.46
Reference-camera orientation (degrees/frame)	0.27	0.12	0.92

Statistics pool valid adjacent-frame transitions within the same world segment. For each wrist measurement, the larger valid left/right change is used at each step. Wrist position is world-coordinate joint 0; orientation is the MANO root rotation transformed into world coordinates, with changes measured as geodesic rotation angles. The counts are 139,187 wrist-translation transitions, 139,165 wrist-orientation transitions, and 139,777 camera transitions. Missing or invalid transitions are excluded, not assigned zero motion. These are motion statistics, including estimation variation, not ground-truth accuracy measurements.

Coordinates and movement

Camera coordinates follow the OpenCV convention: +X right, +Y down, +Z forward. World positions are in meters. Each episode contains one world-coordinate segment: 45 episodes use gravity-aligned +Z-up coordinates, and 14 use the initial camera-axis orientation. Independent episode world frames are not mutually aligned; the convention is recorded per sample.

The median per-episode maximum camera displacement from its starting position is 0.14 m, and the overall maximum is 1.41 m. In 54 of 59 episodes, displacement remains below 0.5 m.

The supplied 2D keypoints are projections of the supplied 3D trajectories using the corresponding calibration and poses. This expresses 2D–3D geometric consistency, not an independently measured image-localization error.

Browse and use

Download the repository with its directory structure intact, then open index.html locally. Select a sample to view its camera video, trajectory visualization, and timed subtasks. HTML previews on hosting services may not run the offline viewer.

• README.html: detailed usage and matching all-episode statistics.
• FIELDS.html: field and coordinate definitions.
• catalog.json: sample IDs, titles, frame counts, durations, and paths.
• field_schema.json: Parquet field types.
• read_sample.py: minimal Python reader.
• samples/sample_NNN/: individual episode data.

samples/sample_001/
  meta/                         Episode metadata and timed subtasks
  data/chunk-000/                Per-frame camera-coordinate data
  videos/chunk-000/              Six camera videos
  renders/                      Trajectory overlay and world-coordinate animation
  support/                      World trajectories, camera poses, calibration and timing

With pyarrow installed, run:

python3 read_sample.py samples/sample_001