LIDAR-GT/HUGG_ARIA_PINHOLE

HUGG ARIA Pinhole

This is a derived, frame-aligned pinhole version of the public `LIDAR-GT/HUGG_ARIA` dataset.

Each RGB frame is generated with the official HOT3D/Project Aria calibration path:

1. 1.Read stream 214-1 from recording.vrs at its exact TIME_CODE timestamp.
2. 2.Query the per-frame online FISHEYE624 calibration.
3. 3.Query the per-frame online LINEAR calibration.
4. 4.Warp the source image with projectaria_tools.core.calibration.distort_by_calibration.

This is pixel-equivalent to AriaDataProvider.get_undistorted_image() at the same timestamp. Every sequence is validated on sampled frames, and the encoded video is decoded again to verify its frame count.

Sequence layout

<SEQUENCE>/
  rgb_214_1_pinhole.mp4
  frame_timestamps_214_1.csv
  frame_pinhole_calibration_214_1.jsonl
  headset_trajectory.csv
  mano_hand_pose_trajectory.jsonl
  masks/*.csv
  source_metadata.json
  metadata.json
  _SUCCESS.json

frame_pinhole_calibration_214_1.jsonl stores the exact target intrinsics and T_device_camera used for every encoded frame. Gaussian/MANO rendering should use the calibration record with the same frame_index and timestamp_ns.

No learned hand detector boxes are generated. A hand box can be obtained from the extrema of valid pixels in a Gaussian/MANO alpha render, which keeps the box synchronized with the rendered hand prior.

Video encoding

Videos are H.264, 1408x1408, 30 FPS, yuv420p, NVENC CQ 20, with a one-second GOP. Encoding is lossy; timestamps and camera calibration are lossless text sidecars.

License

This derived dataset remains subject to the licenses and terms distributed with the source HOT3D/HUGG_ARIA data. Per-sequence license files are retained when available.