Frame-wise stereo geometry registration pilot fails drive-level recovery¶
Frozen at bd9709ee, evaluated 2026-10-30. See
protocol,
executable plan,
all 108 fits and
schema.
Reject adoption of these two native ICP pilots on the prepared stereo geometry. Neither method recovers all three frames of any perturbed drive/start. No confidence, distance, voxel or iteration threshold was retuned after the results.
| Outcome | Point-to-point | Point-to-plane |
|---|---|---|
| Perturbed frame recovery | 1/36 | 0/36 |
| Perturbed drive/start recovery (all three frames) | 0/12 | 0/12 |
| Reference-start frame controls | 6/18 | 2/18 |
| Reference-start drive controls (all three frames) | 1/6 | 0/6 |
| Perturbed numerical convergence | 8/36 | 9/36 |
| Reference numerical convergence | 6/18 | 3/18 |
| Perturbed median rotation error | .8828 deg | .8221 deg |
| Perturbed median translation error | .1700 m | .3097 m |
| Reference median rotation drift | .2223 deg | .2633 deg |
| Reference median translation drift | .0938 m | .1688 m |
The remaining statuses are max_iterations: perturbed 28/36 and 27/36; reference 12/18 and 15/18. Ignoring numerical status, five perturbed frames per method meet the geometric accuracy gates; they are not all accepted recoveries. Reference geometry alone passes in 17/18 point-to-point and 11/18 point-to-plane cases, also not substituted for the declared convergence requirement. This exposes a termination limitation as well as inaccurate converged cases; no iteration-cap extension is applied after observing it.
All three frames of each recording were independently paired. Reference poses were used only in the labeled controls and scoring. No cross-time cloud matching, pose averaging, reference-dependent filtering or best-frame selection occurred. Historical joint MI6/12 uses a three-frame estimator and remains contextual; it is not a matched numerical comparator for this single-frame pilot.
Evidence and verification¶
Inputs are the previously prepared 18 arrays, with 25536–42927 source points and 38351–71126 target points. Source/target cardinality is not overlap. The report retains correspondence counts, train/holdout counts, error diagnostics, conditions, iteration history and resolved options. Complete results including point split IDs are gzip artifacts under /workspace/work/stereo-registration-pilot-20261030; paths and SHA-256 values are retained in every case. These local runtime artifacts must be copied or regenerated when moving the workspace.
Total solver time including diagnostics: point-to-point 257.56 seconds, point-to-plane 293.13 seconds. Input loading and report compression are excluded; this is not a controlled performance benchmark. Point-to-point multi-start inspection was explicitly disabled per the prior protocol; all other settings and spatial holdout policy remain as declared.
Thirteen focused tests, repository Ruff and strict documentation build pass. Validation checks report/schema, 108-case completion, all frozen source and input hashes, every detailed-result artifact hash, split counts, and disjoint training/ holdout point IDs. Native solver and production defaults are unchanged. No fresh recording, trained checkpoint, external code execution or SOTA result is added.
Next decision¶
These inputs do not yet establish a dependable geometric correspondence provider. Before another registration variant, quantify whether each prepared stereo target actually provides accurate local geometry around the LiDAR surfaces: at reference pose, measure nearest-target distances and coverage, and compare stereo target normals with local LiDAR normals where supported. Keep this as offline measurement diagnosis, not truth-filtered fitting. Include all 18 frames and retain missing/ weak neighborhoods. This separates insufficient overlap/noisy surfaces from the solver's iteration limit before spending effort on another optimizer or adding a new dependency. Preserve the existing full-cloud MI baseline and all failures.