Skip to content

Disjoint geometric validation results

Protocol frozen at 3249838a. All 40 candidate checks (120 frame checks) completed without rerunning calibration. All validation clouds are disjoint from fitting points and retain camera occlusions.

Condition Candidate Geometry pass Pose gate pass Worst frame median m Worst frame p90 m
exact_z_visible reference_control True True 0.0039 0.0211
exact_z_visible wrong_control False False 0.5000 11.6051
exact_z_visible baseline_1 False False 0.0817 0.2105
exact_z_visible continuation_1 True True 0.0329 0.1085
exact_z_visible baseline_-1 False False 0.0615 0.1823
exact_z_visible continuation_-1 False False 0.0615 0.1823
exact_z_visible baseline_0 True True 0.0047 0.0229
exact_z_visible continuation_0 True True 0.0047 0.0229
exact_z_all reference_control True True 0.0039 0.0211
exact_z_all wrong_control False False 0.5000 11.6051
exact_z_all baseline_1 False False 0.1209 0.2622
exact_z_all continuation_1 True True 0.0371 0.1200
exact_z_all baseline_-1 False False 0.0692 0.2331
exact_z_all continuation_-1 False False 0.0594 0.1849
exact_z_all baseline_0 True True 0.0045 0.0213
exact_z_all continuation_0 True False 0.0475 0.1546
exact_range_visible reference_control True True 0.0039 0.0211
exact_range_visible wrong_control False False 0.5000 11.6051
exact_range_visible baseline_1 True True 0.0307 0.0929
exact_range_visible continuation_1 True True 0.0168 0.0480
exact_range_visible baseline_-1 False False 0.2006 0.2966
exact_range_visible continuation_-1 True False 0.0358 0.1020
exact_range_visible baseline_0 True True 0.0052 0.0300
exact_range_visible continuation_0 True True 0.0052 0.0300
biased_z_visible reference_control False True 0.2502 0.2613
biased_z_visible wrong_control False False 0.3361 11.7582
biased_z_visible baseline_1 False False 0.3250 0.4600
biased_z_visible continuation_1 False True 0.2269 0.2789
biased_z_visible baseline_-1 False False 0.2630 0.3935
biased_z_visible continuation_-1 False False 0.2630 0.3935
biased_z_visible baseline_0 False True 0.2502 0.2630
biased_z_visible continuation_0 False True 0.2502 0.2630
holes_z_visible reference_control True True 0.0039 0.0207
holes_z_visible wrong_control False False 0.5383 11.7145
holes_z_visible baseline_1 False False 0.0763 0.1985
holes_z_visible continuation_1 True True 0.0283 0.0696
holes_z_visible baseline_-1 False False 0.0695 0.2014
holes_z_visible continuation_-1 False False 0.0562 0.1558
holes_z_visible baseline_0 True True 0.0078 0.0262
holes_z_visible continuation_0 True True 0.0078 0.0262

The proposed gate fails its key control

The all-return continuation from the true pose, already known to drift by 0.8560 degrees / 0.1427 m, passes the new validation. Its worst-frame absolute median is about 0.048 m and p90 0.155 m. The exact-range negative-seed continuation also passes despite 0.5731-degree rotation error (median 0.036 m, p90 0.102 m). Disjoint points and small aggregate depth residuals therefore do not establish full-pose accuracy even in this noiseless analytic scene.

Across the 40 checks, 20 meet the pose gates and 20 do not. The validator accepts 16 correct poses and 2 incorrect poses, rejects 18 incorrect poses and 4 correct poses. All four correct-pose rejections belong to the +0.25 m biased camera-depth condition, including its true-pose control. Thus a failed consistency check cannot distinguish extrinsic error from measurement bias. Every deliberately large wrong-pose control is rejected, but that easy success does not rescue the nearby false accepts. These counts are descriptive, not independent statistical rates: candidates repeat across initializations and conditions in one shared scene.

The same depth maps and scene geometry underpin optimization and validation. Disjoint sampling prevents point reuse but not shared systematic errors or weak constraints. Depth median/p90 statistics are not rotation/translation uncertainty bounds. Behind/in-front residual counts remain in the report; they are diagnostic signs of disagreement, not proven occlusion labels. No residuals were silently removed to obtain passing scores.

Decision

Do not promote this validator, relax pose gates or tune its .05/.20 m thresholds to reject the two observed failures. In particular, do not claim the continuation optimizer is safe because it passed a geometric holdout. The declared drift check failed. No real-data rerun is warranted under that claim.

The next substantive algorithmic test should compare directly fitting metric geometry against maximizing MI, on the same original fitting samples and fixed initial poses, with these holdout samples reserved for diagnostics. Keep depth bias and camera occlusion conditions; they expose different failure modes. Freeze a bounded robust metric estimator and its budget before testing, retain the MI baseline and full-pose labels, and do not let held-out thresholds select fitting parameters. This is a change of calibration objective, not another acceptance threshold variant. Any synthetic improvement still needs broader independent geometry and real-world evidence before promotion.

Evidence and schema include all support counts and metrics. Tests cover depth units, deliberate metric error, missing support and spatial disjointness. Six focused tests, repository Ruff, strict docs build, frozen-hash and report/population checks pass. Production code/defaults and previous experimental verdicts remain unchanged. No SOTA claim.