Known Issues & Data Quirks

A short list of quirks to handle when loading RealSim-CP. The visualizer already tolerates all of these; if you write your own loader, account for them too.

Point-cloud gaps (Daiba daytime variants)

• sunny_1 has LiDAR folders for only 8 of the 12 image agents, and each folder currently contains only one .pcd file instead of one per frame.
• sunny_2 has no point_clouds/ directory at all — images and the label JSON only.

The label JSON in these variants still references the missing point-cloud URIs, so a loader should skip missing files gracefully (degrade to "no point cloud" for that agent/frame) rather than erroring.

Want full LiDAR coverage?

Use a night variant (night_1 / night_2) — these have complete per-frame LiDAR for all agents.

Coordinate / projection quirks

• Quaternion order in cuboid vectors is xyzw ([x, y, z, qx, qy, qz, qw, L, W, H]), not wxyz.
• Vehicle agent pose ≠ body center. vehicle_*_local sits near the rear-axle ground point. Use the matching labeled object cuboid to draw the vehicle body. See Label format.
• Camera image axis: projection uses +X forward, +Y image-left, +Z up, so flip the horizontal axis when drawing to pixels.
• Distortion vs. pinhole: some rendered images align better with a pinhole model than with the full k1…k6 distortion. Make distortion optional in your projection.
• Per-frame transforms exist for moving agents (vehicles); static RSU frames may instead rely on coordinate_systems[rsu_*_local].pose_wrt_parent.

Reporting problems

Found something off in the data? Please open an issue on GitHub.