Fixed reverse consistency does not resolve the remaining surface failures¶
Frozen d6bcae4e, evaluated2026-10-30. See
protocol,
plan,
all18 conditional diagnoses
and schema.
Do not use the1px reverse mask as a remedy for the three failed surface frames. All three remain below80% normal agreement after cycle selection. This is a new conditional diagnosis of predictive surfaces, not a rerun or revised verdict of the earlier translation-refinement experiment.
| Previously failed frame | Before cycle agreement | Cycle-pass agreement | Selected pixels retained |
|---|---|---|---|
| 0005/frame38 | 42.70% | 45.71% | 93.78% |
| 0005/frame114 | 76.28% | 75.50% | 95.83% |
| 0009/frame111 | 76.43% | 78.83% | 93.69% |
Normal agreement improves in13/18 frames and worsens in five. Across-frame median agreement is88.16% before and87.95% in the cycle-pass subset: paired improvement counts and medians summarize different aspects and must both be disclosed. Median distance-failure fraction decreases9.14→8.57%. Median selected-domain retention changes24.77→24.37%; individual retained fractions remain above12.39% of the domain. These modest changes do not establish calibration improvement.
Rejected-or-reverse-missing selected points have median normal failure30.04%, versus12.05% among cycle-pass points, and median distance failure27.16% versus8.57%. This indicates useful association with some errors but does not make accepted points reliable: reverse-consistent surfaces can still disagree strongly with LiDAR. The removed group combines inconsistency and absent reverse support; it must not be interpreted as a labeled set of incorrect correspondences.
Reproduction and scope¶
All18 forward/restored SGBM depth arrays exactly reproduce the archived joint stereo outputs. All12 previously generated cycle-filtered depths also reproduce array-for-array. The other six frames belong to the already-spent0013/0015 drives. Threshold, matcher settings, reverse direction, rectification and restoration are unchanged. Candidate plane estimates were not refitted on masked depth; the test only partitions the already selected points using independent reverse processing.
Five focused tests, repository Ruff and strict documentation build pass. Validation checks schema,18-frame completion, all source/input hashes, partition count totals, normal-support totals and saved filtered-depth hashes. No new optimizer runs, source masks, ICP adoption, production defaults or calibration verdicts change. Filtered arrays remain under /workspace/work/surface-cycle-20261030, with artifact paths and hashes in the report. Reference distance/normal maps are scoring-only.
Next direction¶
Close reverse-threshold/filter adjustments. Review the feasibility of temporal consistency using independent camera motion and overlapping observed surfaces. The three archived frames per drive are separated in time, so overlap and motion availability must be audited before prescribing a temporal filter. Identify pose source, camera/IMU timing/extrinsics, scale and frame conventions; do not estimate camera motion from reference camera/LiDAR extrinsics or silently apply a truth transform in selection. If independent motion or overlap is inadequate, report that limitation rather than manufacturing a successful temporal test. Freeze the distinct measurement comparison before any new fit or threshold selection. No fresh holdouts or SOTA claim are authorized by this spent-data diagnostic.