3D point cloud annotation is unforgiving. A bounding cuboid that is 30 cm too generous does not look wrong to a reviewer scrolling frames — but it systematically teaches a perception model that vehicles are larger than they are.
1. Over-sized cuboids on sparse returns
Distant objects return a handful of points. Annotators padding boxes "to be safe" inflate object dimensions exactly where the model has the least evidence. We enforce dimension priors per object class and flag statistical outliers automatically.
2. Identity switches across frames
Temporal tracking is where most datasets rot. Two vehicles crossing paths swap IDs, and every downstream motion-prediction label inherits the error. Our tooling replays trajectories and flags physically implausible jumps.
3. Ignoring calibration drift
Camera–LiDAR projections that were perfect in week one drift after re-mounts. We verify extrinsics on every delivery batch, not once per project.
4. Ground-plane leakage
Points from road surface bleeding into low cuboids corrupt height estimates. Class-conditional ground filtering keeps boxes tight.
5. Unlogged edge-case policy
Trailers, mirrors, open doors — whatever the policy is, it must be written, versioned, and applied consistently. Ambiguity that lives in annotators' heads becomes noise in the dataset.
