Camera–LiDAR: restart only from a better training candidate¶
This paired development experiment tests the search deficits identified in the geometry failure analysis. It compares the existing neighbor-geometry objective against the same objective with bounded training-only baseline restarts. It is not another objective change.
Fixed restart rule¶
Start with the existing training-only baseline fit and its two nuisance-refitted signed controls. Clip each control seed to the existing rotation bounds and score it on training frames. Refine only seeds strictly better than the original baseline, releasing all three rotation axes. Run at most two refinements, using the existing search resolution. Keep the highest training score, retaining the original baseline on ties. If the selected baseline changes, recompute both controls once. Never iterate until a control passes.
The gate compares both seeds with the original baseline score, not with the result of the first restart. Held-out frames are consulted only after all poses are fixed. Reference calibration is never used as a seed or selector. This gate also avoids spending extra search on the targeted 0002 pitch fits that showed no better training candidate in the previous analysis.
Source 308885cd was frozen in an isolated worktree before running all fourteen
reacquired spent inputs. The arm called original in the report is unrestarted
neighbor geometry, not the old scalar objective. Its saved axis reports must
match the preceding geometry experiment within absolute 1e-9 before the restarted
arm runs. All input hashes, folds, errors and failures are retained. Training-only
initialization, nuisance multi-starts, translation, control sizes and the 2.5 t
threshold are unchanged.
The focused 26 tests pass. They cover strict improvement, skipping ties and worse seeds, clipping before gating, comparison against the original score for both seeds, and the existing training/held-out separation and geometry invariants.
Reproduce from the frozen source:
OPENBLAS_NUM_THREADS=1 PYTHONPATH=src:. python -m tools.kitti_camera_lidar_boundary_batch \
/path/to/capture --output /path/to/new/output --geometry-restart
These are 42 correlated axis controls from two development drives, not independent calibration evaluations. Full rotation norms are reference disagreements of training-fold fits, not ground-truth accuracy. Concurrent-worker timings are not a repeated runtime benchmark. No result alone authorizes a default change or a SOTA claim.
Results and decision¶
The two lost detections are recovered, with no detected-control losses in this batch. Retain experimentally; do not promote to a native default. All fourteen unrestarted-arm parity checks pass within absolute 1e-9, and all paired inputs complete.
| Diagnostic | Neighbor geometry | With deficit-gated restarts |
|---|---|---|
| Detected controls, 0005 | 19/21 | 21/21 |
| Detected controls, 0002 | 0/21 | 0/21 |
| Total detected controls | 19/42 | 21/42 |
| Worst full rotation error among detected folds | 0.296 deg | 0.296 deg |
| Worst full rotation error across all axis folds | 12.067 deg | 13.568 deg |
| Target 0002 pitch +3 control t | 1.53886 | 1.53886 |
| Target pitch +3 detected | No | No |
| Sum of profile wall times (concurrent workers) | 940.0 s | 1000.4 s |
The recovered controls are roll and yaw on 0005 pitch +3. Roll's t statistic increases from 1.99254 to 2.63272; yaw's increases from 2.20727 to 3.79188. The previously failing fold rotation errors decrease from 2.002 to 0.206 deg and from 2.101 to 0.274 deg respectively. The roll result is only modestly above the unchanged 2.5 threshold. All nineteen previously detected controls remain detected; no additional control is detected on 0002. The worst detected-fold rotation disagreement does not increase. This does not prove a general false-positive rate or safety: all-fold disagreement worsens, and the recordings are the same repeatedly used development data.
Fifty of 252 axis-fold baselines change: the two diagnosed 0005 failures and 48 folds on 0002. Thus the strict training gate is not a universal accuracy safeguard. The difficult 0002 data still supports distant high-scoring basins, even though they do not become detected controls in this experiment. The targeted pitch-axis baseline and its six control pairs remain unchanged, consistent with the previous finding that they contained no better training control seeds.
This separates the recovered search regressions from the unresolved correspondence ambiguity. Keep the bounded restart alongside the geometry method as a development candidate. Future work should focus on the remaining 0002 failures, not lower the detection threshold or infer SOTA performance from 0005 reaching 21/21.
The complete report, summary including all changed baselines and hash manifest retain all fourteen pairs, full rotation vectors, control outcomes and source/input digests.