code/docs/PROTOCOL.md

KITTI-C2L v1 evaluation protocol

Use the exact files and hashes in release/v1/manifest.json. For the recommended benchmark, remove sequence 07 from the source training file and use it only for validation. Test on sequences 09 and 10. All five stages of a frame, and every window containing it, belong to the same partition.

Targets

Native evaluation uses the frozen sensor perturbations in the Parquet files. Construct T_gt and T_init in the same unit system. The UniCal adapter target is D = T_init @ inverse(T_gt); it predicts decalibration, so the correction applied to the input extrinsic is its inverse. Do not regress the raw sampled camera/LiDAR translations as if either one were the composed relative target.

The secondary usual-relative control retains the source frames and GT but replaces the input error by one camera-relative transform drawn uniformly at ±1 degree and ±10 cm per axis, so T_init = D_relative @ T_gt. It is not the frozen native dataset. Our existing control implementation uses UniCal's ErrorGenerator and, for validation/test, a Torch generator seeded with sample_seed % 2**63 (relative_seed zero). The exact sampler and rotation conversion live in the pinned UniCal evaluation source; substituting NumPy sampling will not reproduce those targets. Online training draws vary; the aligned experiment changes both training and validation perturbation policies.

Metrics

For N targets D_i = (R_i, t_i) and predictions (Rhat_i, that_i), report:

• Translation MAE: sum_i sum_j abs(that_ij - t_ij) / (3*N), in centimetres. Multiply metre-based errors by 100. This is not Euclidean translation norm.
• Rotation: mean geodesic SO(3) angle between Rhat_i and R_i, in degrees. For valid rotations it is acos(clip((trace(Rhat_i.T @ R_i)-1)/2,-1,1)). Our experiment evaluator uses SciPy Rotation normalization for small numerical deviations from orthogonality. Euler-angle MAE is a different metric and cannot be compared directly with these numbers.

Include identity correction, per-drive results and the aggregate. Report sample counts, seeds, selected checkpoint epochs, image size, point sampling, normalization, precision, training updates and encoder-frame exposure. Select checkpoints on validation, never on the test values reported here.

The existing paired comparisons resample training seeds and shared temporal blocks of 30 anchors, using 5,000 bootstrap draws and 95% intervals. Their 10% noninferiority screen is a study decision rule, not a universal definition of dataset validity. Dependence remains within this single KITTI rig and the two held-out drives; intervals do not establish generalization to other rigs.

Temporal methods

Use the windows configuration when assuming a fixed error over consecutive frames. Reset state at each sequence, stage and window boundary. Final windows can be shorter than 30 frames. Report the minimum context and anchor-selection policy. Our three-frame comparison uses 2,604 common anchors, versus 2,792 for single-frame evaluation. Compare methods on identical anchors; a smaller anchor set changes the benchmark. Disclose whether fusion is inference-only or trained, and match encoder exposure when claiming training efficiency.

Geometry version

The frozen v1 camera-2 calibration uses rectification plus the x-only stereo baseline shift. The full K^-1 @ P[:,3] baseline differs by up to about 0.62 cm on these dates. The source audit validates agreement with the frozen producer convention, not equivalence to every alternative calibration construction. Keep v1 unchanged so results remain comparable. Future geometry changes need new hashes, a new version and updated baselines.