Datasets:

---

nvidia

/

PhysicalAI-Autonomous-Vehicles

like 1.01k

Follow

NVIDIA 68k

Summary

This PR updates the offline LiDAR-camera calibration for 115 clips (55 recording sessions; clips from the same session share one calibration) whose published calibration was misaligned (13-28 degrees on the worst camera; in several cases the projection center was outside the image).

What changes

• calibration/camera_intrinsics.offline/*.parquet (101 chunks): model_parameters replaced for all 7 cameras of each affected clip (principal point, FTheta polynomials, max angle, linear terms; windshield model added where the lens has one). model_type and ego_mask_image_png are unchanged.
• calibration/sensor_extrinsics.offline/*.parquet (101 chunks): qx, qy, qz, qw, x, y, z replaced for the 7 camera rows of each affected clip. lidar_top_360fov rows are not modified.
• README.md: new 26.09 row in Version History.

What does not change

• Every other row in the 101 chunk files (other clips, LiDAR, radar) is byte-for-byte equal to the base file.
• Online calibration (camera_intrinsics, sensor_extrinsics), vehicle_dimensions, lidar_intrinsics.offline, metadata, clip_index.parquet, features.csv, video, LiDAR, radar and labels are untouched.
• Parquet schema, index order, dtypes, compression and Arrow metadata are identical to the base files.

Validation

• Every patched camera row passes the FTheta model_parameters schema check; every patched extrinsics row is finite with quaternion norm and rotation determinant within 1e-6 of 1.
• The published lidar_top_360fov extrinsics of every affected clip agree with the calibration source within 1e-6, confirming both refer to the same rig.
• Before/after LiDAR-to-camera projection overlays were reviewed visually for the affected clips.
• Affected clips by split: train 61 / val 26 / test 28.

After merge

Tag the merge commit as 26.09 to match the Version History row.