rooty2020/seoyeon

Multi-view video + 3D point tracks (RB-Y1 x10, ORH C001)

rby1/
  task_instructions.jsonl        caption per episode (episode_index 0-9)
  episode_000 ... episode_009/
    ego.mp4, exo.mp4             2 views, 1280x720, 15 fps, H.264 intra-only
    robot.npz                    points (T,512,3) float32  arm surface points
    hand.npz                     points (T,256,3) float32  WujiHand2 hand points
    meta.json                    intrinsics, dist, T_ego_base / T_exo_base, fps, n_frames
    action.parquet               actions @15 Hz (not needed for video/tracks)
orh4d_C001/
  videos/<serial>.mp4            47 views, 2048x1536, 15 fps, HEVC, 150 frames (10 s)
  cameras.json                   calibration of the 47 views + rawFrameID per frame
  tracks.npz                     3D tracks, 150 frames x 256 points
  tracks.json                    how the tracks were made
  preview_tracks_4d.mp4          quick-look render of the tracks

Frame t of every video corresponds to frame t of the tracks in the same folder.

RB-Y1 (rby1/)

Source: HF kdh8156/RB-Y1_WujiHand2_teleop_ego_exo_100 and its point-track re-export rooty2020/RB-Y1_WujiHand2_teleop_ego_exo_100_pointtracks, episodes 0-9 (171-227 frames each).

• Tracks are ground truth from forward kinematics (robot MJCF + joint positions), so point i is the same physical point on the robot in every frame. No occlusion or visibility flag is stored; if you need one, compute in-frustum visibility yourself (z > 0 and inside the image).
• Points are in the robot base frame, not the camera frame. Project to a view with meta.json:

  T = np.array(meta["cameras"]["exo"]["T_exo_base"])       # 4x4, base -> camera
  K = np.array(meta["cameras"]["exo"]["intrinsics"])
  p = pts @ T[:3, :3].T + T[:3, 3]                          # (N,3) in camera
  uv = (p @ K.T)[:, :2] / p[:, 2:3]
  

  Use ego / T_ego_base for the ego view.
• Both cameras are static in the base frame: the per-frame ego pose in meta.json has zero variance, so the static T_<cam>_base is correct for every frame.
• The exo projection is exact. The ego projection has a small residual toward the far left because the source dataset's ego extrinsic was set by hand.

ORH C001 (orh4d_C001/)

Source: HF Sulwon/anonymous_dataset_lih_orh_0920, clip C001. It is a 47-camera rig (camera 23012639 is excluded because it has no frames). The videos are the first 150 frames (rawFrameID 1752..2050, stride 2, i.e. 30 -> 15 fps) of the 15 s source clip. They were cut with a stream copy, so they were not re-encoded.

• cameras.json → cameras[<serial>]: K_original, dist_params (OpenCV k1,k2,p1,p2,k3), T_cam_from_world (w2c 4x4), T_world_from_camera, camera_center_world, K_undistort. OpenCV axes, metres. The videos are the original distorted images, so project with cv2.projectPoints(xyz, rvec(R), t, K_original, dist_params).
• tracks.npz:

  key	shape	meaning
  xyz	(150,256,3)	3D track positions, world metres (same frame as cameras.json)
  vis	(150,256)	tracker visibility probability in [0,1]
  query	(256,3)	the frame-0 query points
  uv	(150,256,2)	tracks projected into camera 22684751, normalised [0,1] image coords
  uv_ok	(150,256)	that projection is in-frame and not behind the surface

• These tracks are predictions, not ground truth. The 4D reconstruction (per-frame point clouds from orh_4d_reconstruction) does not keep point identity across frames. The tracks were therefore predicted with MVTracker, using 4 views (22684751, 22645022, 22684259,
  22640993. resized to 512x384 and depth rendered from each frame's reconstructed cloud. The 256 queries are farthest-point samples on the moving subject at t=0. Median per-frame step is 8.2 mm; about 61% of (frame, point) pairs are tracker-visible.