Camera–LiDAR: neighboring points on the same image contour

Decision: do not integrate this proposal into pose search. On the saved pitch failure, a wrong control still beats the dataset reference in five of six folds. This screen avoids a full optimization batch for a proposal that fails to reject known incorrect poses. It establishes no new accuracy or SOTA claim.

Fixed hypothesis

Independently selecting nearby image ridges permits points to switch between unrelated boundaries. Test whether requiring nearby depth-boundary points to match the same image contour suppresses those substitutions.

Before scoring, source 40ae85d4 fixes a mutual four-nearest-neighbor graph over prepared 3D depth-boundary points, with a 0.75 m maximum link length. Label the existing image ridge pixels with eight-connected components; propagate those labels to the nearest pixel within six pixels. At each saved pose, multiply each point's existing nearest-ridge reward by the fraction of its fixed neighbors matching the same nonzero component. A point with no neighbors keeps its original reward. Unknown matches count as disagreements for linked points. Preserve all point weights and the original score denominator.

This is a local consistency test, not semantic correspondence. Image components can merge different physical boundaries or fragment one boundary. Three-dimensional neighbors can also cross surfaces. Rounded pixels supply component identity; the existing reward continues to use bilinear interpolation.

The screen uses every saved reference, baseline and signed nuisance-refitted control from all six folds and eleven prepared frames of 0002 pitch +3. There is no new fitting, frame selection or threshold tuning. All original per-frame scores and pooled fold scores reproduce within absolute 1e-9. Four focused tests cover graph construction, partial agreement, unknown labels, distance limits, isolated-point fallback and exact original contribution accounting.

The decision rule was fixed in the source before execution: if a saved control still beats the reference in at least five folds, do not integrate. Passing this screen would only justify a full paired development experiment, not adoption.

Results

The graph covers 6,887 of 7,401 prepared points (93.05%). Its failure is therefore not explained by most points receiving no constraint.

Diagnostic	Nearest ridge	With contour consistency
Folds where a saved control beats the reference	5/6	5/6
Folds where a saved control beats the fitted baseline	4/6	4/6
Frames where a saved control beats the reference	10/11	8/11

The following scores retain both controls, including failures:

Fold	Reference	Baseline	Minus-pitch control	Plus-pitch control
0	0.3110	0.2410	0.2035	0.2302
1	0.2795	0.3323	0.3066	0.2984
2	0.3529	0.3562	0.3737	0.3636
3	0.3283	0.3329	0.3206	0.3439
4	0.3405	0.3124	0.2977	0.3736
5	0.2903	0.2957	0.2421	0.3014

The penalty narrows several wrong-pose score advantages but does not reverse the pooled preference in enough folds. In fold 3 the preferred signed control changes; the plus-pitch control still has 4.478 deg full rotation disagreement and beats the reference. The retained scores show that merely sharing a connected contour is not sufficient to establish correspondence identity.

The next structural hypothesis should check the geometry of neighboring matches (direction, separation and collapse), not only their component labels. That is untested here. Do not claim that all structural matching methods have failed, or that tuning these constants would produce independent validation.

Complete scores and graph support and the hash manifest retain all poses and outcomes. Run tools/kitti_camera_lidar_contour_coherence.py with the reacquired snapshot, the previous pitch-match diagnostic receipt and a new --output directory. Native estimator defaults and earlier audit verdicts remain unchanged.