Full clouds restore MI recovery but do not rescue the metric objective¶
Frozen at 07398b48 before evaluation on 2026-10-30. See
protocol,
plan,
all 54 fits and
schema.
Retain the existing MI path; reject adoption of the current metric candidate. Full-cloud MI exactly reproduces all 12 historical SGBM result records, including traces, transforms and metrics, recovering 6/12 perturbed starts. Both metric methods still recover 0/12. File-order thinning harmed MI, but does not explain away the metric method's failure on these real recordings.
| Full-cloud method | Perturbed recovery | Reference-start controls | Perturbed convergence/support | Perturbed median rotation / translation |
|---|---|---|---|---|
| MI | 6/12 | 5/6 | 12/12 | .3096 deg / .2190 m |
| Bilinear metric | 0/12 | 1/6 | 11/12 | 1.3615 deg / .2611 m |
| Edge + scalar rescue | 0/12 | 0/6 | 11/12 | 1.4104 deg / .7032 m |
All six reference-start fits terminate successfully for each method, yet metric outputs frequently drift outside the pose gates. Median reference drift is MI .1675 deg/.1531 m, bilinear .7662 deg/.2356 m, edge+rescue .7624 deg/.2434 m. Metric convergence/support successes also pass the local rank gate; neither termination nor local rank certifies accuracy. The nested edge-only baseline has identical poses to the rescue output. Sixteen scalar triggers/adoptions lower cost without changing any pose; this does not produce a pose recovery gain.
Compared with the selected-point experiments, MI perturbed recovery improves 1/12→6/12, bilinear remains 0/12 and edge+rescue remains 0/12. Reference-start scores change MI 5/6→5/6, bilinear 1/6→1/6 and edge+rescue 3/6→0/6. The complete input contains 107739–124044 points per frame; every point is passed through, without a new visibility mask, sampling or revised frontend. Algorithms retain their own projection and correspondence rules, so this comparison of full methods does not isolate the objective formula from the search algorithm.
Cost preference survives full input¶
At equal fitted bias, fitted cost is below reference cost in 18/18 bilinear cases, 17 with inaccurate poses. On identical common-support point indices the same 18/18 and 17 inaccurate cases remain. Edge+rescue has lower total cost in 17/18, all inaccurate, and lower common-support cost in 15/18, all inaccurate. Thus support membership changes alone do not account for the wrong-pose cost preference. These are fixed-state probes, not globally optimized reference-bias comparisons.
Reference-zero supported populations are 41918–46841 points per three-frame batch, out of 335032–368538 total points. Depending on drive, about 14.8–29.5% of supported signed depth residuals exceed 1 m in magnitude. This error is present at reference poses with both samplers. It warrants studying the measurement-error model; it does not by itself prove which of stereo, occlusion, timing or search causes every failure. Real nuisance-bias accuracy is unknown.
Every full-cloud outcome¶
Each cell gives rotation in degrees / translation in metres / recovery.
| Case | MI | Bilinear | Edge + rescue |
|---|---|---|---|
| 2011_09_26_0005-positive | 0.8773 / 0.2577 / False | 0.7516 / 0.2053 / False | 0.7037 / 0.6965 / False |
| 2011_09_26_0005-reference | 0.1875 / 0.0904 / True | 0.4446 / 0.1630 / True | 0.4700 / 0.2171 / False |
| 2011_09_26_0005-negative | 0.3149 / 0.1803 / True | 1.3356 / 0.1669 / False | 1.3042 / 0.7980 / False |
| 2011_09_28_0002-positive | 0.1835 / 0.0400 / True | 1.3671 / 0.1065 / False | 2.9478 / 1.4668 / False |
| 2011_09_28_0002-reference | 0.1474 / 0.0265 / True | 0.7066 / 0.1723 / False | 0.5408 / 0.1623 / False |
| 2011_09_28_0002-negative | 0.5553 / 0.3536 / False | 1.5672 / 0.4403 / False | 2.2532 / 0.4755 / False |
| 2011_09_26_0009-positive | 0.9009 / 0.2658 / False | 0.5184 / 0.3571 / False | 0.8894 / 0.5641 / False |
| 2011_09_26_0009-reference | 0.1406 / 0.1602 / True | 0.3681 / 0.2027 / False | 0.6607 / 0.2522 / False |
| 2011_09_26_0009-negative | 0.1985 / 0.1250 / True | 2.8131 / 0.2393 / False | 0.9686 / 0.7099 / False |
| 2011_09_26_0011-positive | 0.1371 / 0.0468 / True | 1.0653 / 0.2409 / False | 1.5167 / 0.8016 / False |
| 2011_09_26_0011-reference | 0.0781 / 0.1579 / True | 1.6247 / 0.3347 / False | 0.8792 / 0.2930 / False |
| 2011_09_26_0011-negative | 0.2782 / 0.3197 / False | 3.1093 / 0.3811 / False | 1.6354 / 0.6155 / False |
| 2011_09_26_0013-positive | 0.6802 / 0.3085 / False | 0.7051 / 0.4026 / False | 0.5839 / 0.4333 / False |
| 2011_09_26_0013-reference | 0.6320 / 0.1945 / False | 1.8058 / 0.4621 / False | 0.9784 / 0.2346 / False |
| 2011_09_26_0013-negative | 0.8442 / 0.4727 / False | 2.4919 / 0.2814 / False | 1.7396 / 0.2276 / False |
| 2011_09_26_0015-positive | 0.3044 / 0.1281 / True | 1.3559 / 0.8844 / False | 1.0671 / 1.0793 / False |
| 2011_09_26_0015-reference | 0.2210 / 0.1484 / True | 0.8259 / 0.2685 / False | 0.8642 / 0.5137 / False |
| 2011_09_26_0015-negative | 0.1214 / 0.0904 / True | 2.4066 / 0.2045 / False | 1.9391 / 0.7968 / False |
Validation and decision¶
MI uses 157–421 evaluations; bilinear 97–315 and edge+rescue 105–208 actual residual calls, all below their respective 800 caps. Total solve times including solver diagnostics are 68.63, 177.60 and 188.90 seconds respectively, excluding loading and separate cost probes. MI evaluations and metric calls have different semantics; this is not a rigorous speed benchmark.
Nine focused tests, repository Ruff and strict documentation build pass. Validation checks all frozen source/input hashes, report schema, all 54 cases, full input counts, hard budgets, unchanged rescue poses, cost descent, reproduction of fitted costs and exact historical MI equality for all 12 perturbations. No core or default changes. No new recordings or fresh holdouts were consumed.
Close the sequence of edge-sampler/scalar-rescue tuning. They improve synthetic bias controls but do not provide a viable real-data calibration path. Next build an independently specified stereo uncertainty control: disparity error maps into range-dependent depth error, whereas the current fixed metric residual scale and single additive bias assume a simpler model. First test that hypothesis on known geometry with predeclared disparity noise and occlusions; compare uncertainty-aware residuals to this fixed baseline without fitting noise scales on these failed real cases. Keep MI as the real-data baseline and preserve all truth-start/regression controls. This is a next experiment, not a claim that uncertainty weighting already solves calibration or that MI is SOTA.