CoolFace
Datasetpublic

JimmyH16/cameraCalibrationData

Galaxea R1 Lite distortion-corrected camera sample This is a stationary ROS 2 data-collection test made on 2026-07-26. No chassis, torso, arm, gripper, IK, teleoperation, odometry, or pose-control command was run during the test. Result Stream Resolution Messages Effective bag rate Head left, rectified 1920 x 1536 318 26.1 FPS Head right, rectified 1920 x 1536 320 26.3 FPS Wrist left, unchanged 1280 x 720 364 29.9 FPS Wrist right, unchanged 1280 x… See the full description on the dataset page: https://huggingface.co/datasets/JimmyH16/cameraCalibrationData.

sourceHugging Faceotherupdated 2mo agoView on Hugging Face
0likes15downloads
Dataset Card

Galaxea R1 Lite distortion-corrected camera sample

This is a stationary ROS 2 data-collection test made on 2026-07-26. No chassis, torso, arm, gripper, IK, teleoperation, odometry, or pose-control command was run during the test.

Result

StreamResolutionMessagesEffective bag rate
Head left, rectified1920 x 153631826.1 FPS
Head right, rectified1920 x 153632026.3 FPS
Wrist left, unchanged1280 x 72036429.9 FPS
Wrist right, unchanged1280 x 72036129.7 FPS

The bag duration is 12.1689 seconds. It contains 21,943 original messages plus one /dataset/distortion_fix marker. A byte-for-byte verification found that all 21,943 source messages, timestamps, topic names, and serialized payloads were preserved in the final package.

Correction method

Only the two head-camera streams are changed. Each camera is captured directly from its 1920 x 1536 YUYV V4L2 device, then rectified using an NVIDIA VPI 3.1 CUDA warp generated from the company-provided calibration file:

/opt/galaxea/body/stereo_ros.yaml

Calibration SHA-256: 195ac1482e7d485cdb94a4a76fb42ed6faf706aefad1edb6e5b836b03701f00a

The rectified image is JPEG encoded at quality 90 and published on the same head-camera topic names used by the original recorder. A small C++ ROS 2 publisher avoids Python serialization overhead. Head CameraInfo uses zero distortion coefficients and the official rectified stereo projection matrices. The right projection retains the stereo baseline (P[3] = -66.4200456).

The wrist-camera streams and all non-camera packets are unchanged. The final packaging step does not decode or re-encode any image.

Files

  • —20260726_DISTORTION_FIXED/...mcap: complete ROS 2 MCAP episode.
  • —20260726_DISTORTION_FIXED/metadata.yaml: ROS bag metadata plus correction annotation.
  • —20260726_DISTORTION_FIXED/distortion_fix.json: human-readable correction metadata.
  • —four_camera_20260726_DISTORTION_FIXED.mp4: 1920 x 1080 four-camera preview at 15 FPS.

Why the previous method was slow

The older Python live pipeline subscribed to already-compressed JPEG ROS messages, decoded both full-resolution images, remapped them, encoded them again, and serialized large ROS messages in Python. The correction math was not the main bottleneck. Direct V4L2 capture, CUDA remap, and a C++ publisher remove the expensive ROS/Python path.

Operational notes

The vendor package files were not modified for this solution. The preferred rollback is to stop the Jimmy rectifier and restart the unchanged vendor head camera. The backup snapshot is stored at:

/home/nvidia/jimmy/camera_distortion_backup_20260726_before_company_calibration