pengrui-rio/FEAGINE_HumanEgo_DataSample

Human Ego Hand Trajectory Dataset

中文说明

Overview

This package contains synchronized egocentric manipulation episodes recorded by six cameras. Each episode provides:

• six upright camera videos at 1920 x 1080 and 30 FPS;
• frame-aligned 2D hand observations from all six camera views;
• metric 21-joint 3D hand trajectories in one shared camera reference frame;
• bilingual dense action annotations with temporal segments;
• a synchronized trajectory review video and semantic timeline page.

This sample contains 8 episodes. Using hand-frames with available 2D hand observations as the denominator, the overall 3D hand-trajectory coverage is 98.7% (6090 / 6170), and every episode exceeds 98% coverage. The trajectories can be projected into each original camera view using the provided six-camera calibration. In views where the hand is visible, the projected skeleton remains consistently aligned with the corresponding hand. Exact per-episode coverage values are listed in meta/episode_index.csv.

Review Overlay Provenance

For every hand-frame represented in hand_pose/trajectory_3d.parquet, the corresponding skeleton in review/review.mp4 is generated by projecting the delivered metric 3D trajectory through the six-camera calibration. It is not a direct rendering of hand_pose/observations_2d.parquet.

For the small number of hand-frames without a 3D row, the review video may use an auxiliary 2D observation only to preserve visual continuity. Such a fallback overlay is not part of trajectory_3d.parquet and must not be interpreted as a delivered 3D trajectory. Numerical use should treat trajectory_3d.parquet, coordinate_system.json, and calibration/headset_02_1080p.json as authoritative.

Key Highlights

• High 3D coverage: 98.7% overall, with per-episode coverage ranging from 98.01% to 99.81%.
• Unified six-camera geometry: all 3D trajectories use one metric mid_cam_left reference frame.
• Reproducible projection: complete intrinsics, distortion parameters, and reference-camera extrinsics are provided for all six cameras.
• Complete time alignment: videos, 2D observations, 3D trajectories, and semantic segments share one 30 FPS timeline.

Per-Episode 3D Trajectory Coverage

Coverage is calculated as hand-frames with an available 3D trajectory divided by hand-frames with an available 2D hand observation.

Episode	Video frames	Frames with both hands in 3D	Frames with one hand in 3D	3D trajectory hand-frames	Observed 2D hand-frames	Coverage
episode_000001	228	182	46	410	416	98.56%
episode_000002	228	226	2	454	455	99.78%
episode_000003	255	248	7	503	509	98.82%
episode_000004	303	289	14	592	604	98.01%
episode_000005	525	519	6	1044	1046	99.81%
episode_000006	165	123	42	288	293	98.29%
episode_000007	825	774	51	1599	1630	98.10%
episode_000008	612	588	24	1200	1217	98.60%
Total	3141	2949	192	6090	6170	98.70%

Directory Layout

tasks/
  task_NNNN__english_task_name/
    episode_NNNN__episode_NNNNNN/
      episode_manifest.json
      videos/
        left_cam_left.mp4
        left_cam_right.mp4
        mid_cam_left.mp4
        mid_cam_right.mp4
        right_cam_left.mp4
        right_cam_right.mp4
      hand_pose/
        observations_2d.parquet
        trajectory_3d.parquet
        coordinate_system.json
      semantic/
        annotation.json
      timestamps/
        episode_timebase.json
      review/
        review.mp4
        annotation.html
calibration/
  headset_02_1080p.json
meta/
  episode_index.csv
  task_summary.csv
  delivery_summary.json
  index.html

Open meta/index.html in a browser to browse all episodes. Each episode also contains a standalone review/annotation.html page. All links are package-relative and remain valid after moving the complete dataset directory.

3D Hand Trajectory

Each row in hand_pose/trajectory_3d.parquet represents one hand in one video frame. A row is written only when a 3D trajectory is available; missing trajectories are not represented by NaN rows. Every public keypoints_3d value is finite and has shape 21 x 3.

The two hands are distinguished by the hand field, so one frame can contain up to two rows: left and right. The 8 episodes contain 3,141 video frames in total. All 3,141 frames contain 3D data for at least one hand; 2,949 frames contain 3D data for both hands, and 192 frames contain 3D data for one hand. No coordinates are fabricated for a hand that is outside the view or lacks an available 3D result, and no NaN placeholder row is written. Use (frame_index, hand) as the composite key when constructing a frame-major bimanual tensor.

Missing-hand semantics are defined as follows:

• A hand == "left" row means that a left-hand 3D trajectory is available for that frame; the same rule applies to right.
• If a frame has no left row, no usable left-hand 3D trajectory is available for that frame. It must not be interpreted as a zero vector or as the coordinate origin.
• If observations_2d.parquet still contains a left-hand observation at that frame, the frame has 2D evidence but no available 3D trajectory. If no camera contains a left-hand 2D observation, the left hand is generally absent, outside the field of view, or fully invisible.
• The package contains 2,995 left-hand 3D rows and 3,095 right-hand 3D rows, for a total of 6,090 hand-frames.

Field	Description
frame_index	Zero-based episode video frame index.
timestamp_s	Time in seconds relative to the episode start.
hand	left or right.
keypoints_3d	21 x 3 metric hand-joint coordinates.
reference_camera	Always mid_cam_left in this package.
coordinate_units	Always meters in this package.

The 3D coordinates use the upright mid_cam_left optical frame:

• +X points to image right;
• +Y points to image down;
• +Z points forward from the camera.

The exact 21-joint order and axis definitions are stored in hand_pose/coordinate_system.json.

21-Joint Index Mapping

The left and right hands use the same index definition and are distinguished by the per-row hand field.

Index	Joint name
0	wrist
1	thumb_cmc
2	thumb_mcp
3	thumb_ip
4	thumb_tip
5	index_mcp
6	index_pip
7	index_dip
8	index_tip
9	middle_mcp
10	middle_pip
11	middle_dip
12	middle_tip
13	ring_mcp
14	ring_pip
15	ring_dip
16	ring_tip
17	pinky_mcp
18	pinky_pip
19	pinky_dip
20	pinky_tip

2D Hand Observations

hand_pose/observations_2d.parquet contains frame-aligned 2D observations from the six camera views.

Field	Description
frame_index	Zero-based episode frame index.
timestamp_s	Time in seconds relative to the episode start.
camera	Camera name matching a file under videos/.
hand	left or right.
keypoints_2d	21 x 2 pixel coordinates in the upright 1920 x 1080 image.
confidence	Relative observation-level confidence score; not a calibrated probability.

observations_2d follows this exact contract with the delivered camera videos:

• For a row with camera == "mid_cam_left", keypoints_2d is expressed directly in the pixel plane of videos/mid_cam_left.mp4 in the same episode. The same rule applies to every other camera.
• Each delivered camera video is 1920 x 1080. Coordinates are measured in pixels using the OpenCV image convention: the center of the top-left pixel is (0, 0), +X points right, and +Y points down. Coordinates are not normalized to [0, 1].
• Each raw 3840 x 1080 stereo frame is split into two 1920 x 1080 camera views and then rotated by 180 degrees to produce the delivered upright videos. The 2D keypoints are generated in this delivered pixel plane. No additional resize, crop, or rotation is applied relative to the delivered MP4 files.
• Both the videos and 2D coordinates remain in the distorted camera pixel plane; they are not undistorted. Use distortion_coefficients from the calibration file when projecting 3D points into a video.
• frame_index is zero-based. frame_index = i maps exactly to decoded frame i of videos/{camera}.mp4, and timestamp_s = i / 30. All six videos start at 0 seconds and contain the same number of frames.
• Coordinates are not clamped to the image boundary. When a hand is only partially visible near an image edge, some semantic joints may lie outside [0, 1920) x [0, 1080); users can derive a per-joint in_frame mask from these bounds.

Camera Calibration

calibration/headset_02_1080p.json contains the intrinsics, OpenCV radial-tangential distortion coefficients, image dimensions, and reference-camera extrinsics for all six cameras.

For a 3D point p_ref from trajectory_3d.parquet, the transform convention is:

p_camera_h = T_reference_to_camera[camera] @ [p_ref.x, p_ref.y, p_ref.z, 1]

Project p_camera_h[:3] with the target camera's camera_matrix and distortion_coefficients. Extrinsic translations and 3D trajectories are both expressed in meters.

Time Alignment

All six videos in an episode share the same 30 FPS timeline. timestamps/episode_timebase.json records the episode-relative frame origin and duration. The 2D observations, 3D trajectories, semantic segments, and review video use the same episode-relative timebase.

Semantic Annotation

semantic/annotation.json provides bilingual task summaries and temporally ordered action segments. Each segment includes start and end times, action descriptions, actor roles, interaction phases, contact states, and manipulated objects.

Review Media

review/review.mp4 is a synchronized 1920 x 720, 30 FPS hand-trajectory visualization. For available 3D hand-frames, its displayed 2D skeleton is the six-camera reprojection of the delivered 3D trajectory rather than an independently estimated 2D result. review/annotation.html presents the video together with the semantic timeline. Numerical analysis should use the Parquet files and camera-calibration JSON as the authoritative sources.