KITTI projected-azimuth orientation probe¶
Post hoc development; this proxy is not adopted. This experiment extends the nuisance-profiled controls with image-gradient direction. The existing scalar edge/discontinuity cues have Levinson–Thrun, RSS 2013 as a primary source; the direction penalty below is an independent development hypothesis, not a reproduction of that paper.
Fixed model and scope¶
The diagnostic constructs a polarity-independent image-gradient tensor using the scalar objective's image reduction, Gaussian blur, Sobel derivatives and 99th-percentile gradient normalization. Its trace supplies bounded gradient confidence; no additional pixel-strength threshold is selected.
For a prepared LiDAR point (x, y, z), the azimuth-motion proxy is (-y, x, 0).
Both point and proxy are rotated by the trial transform. The pinhole projection
derivative supplies its image direction, and a tensor cosine-squared term
measures compatibility with the local image gradient. This is scan direction,
not a recovered boundary normal. It need not align with a true image edge,
and deriving it from motion-compensated coordinates adds another approximation.
Only organized frames and undistorted pinhole cameras are supported. Image
gradient orientation is sampled locally; it is not transported along the spill.
The score subtracts raw_response * confidence * (1 - compatibility) from
the original signed response. Thus orientation mismatch cannot turn a negative
background score into a reward: only the nonnegative foreground contribution
is attenuated, while the original background penalty remains. Flat gradients
and zero projected directions are neutral. Point support, denominator and
weights stay fixed, including candidate-outside points entering a trial image.
python -m tools.kitti_camera_lidar_profile_probe SAVED_SNAPSHOT \
--conditional-report PREVIOUS_CONDITIONAL_REPORT.json \
--multi-start --azimuth-orientation --output NEW_DIRECTORY
Orientation and the previously rejected spatial balancing are mutually exclusive ablations. Both default to off. Training frames select nuisance starts; held-out blocks score only fixed poses. Initialization still comes from the pooled conditional fit, so these are not nested validation folds.
Preselected six-input comparison¶
Source c237f78d was committed before launch. The producer fixed baseline,
pitch +3 and roll +3 on each of the two cached spent drives before reading
the new outcomes: six inputs, eighteen axis-control comparisons. These
post hoc probes target the known detection/false-confidence cases. They do
not cover yaw +3, one-degree controls or the full fourteen-input recipe.
Two independent worker processes ran the inputs, with BLAS thread limits requested as one via environment variables. Timing is an execution receipt, not a speed benchmark. All six completed without errors, checking identical snapshot/NPZ hashes, conditional receipts and optimizer settings against their previous scalar multiple-start reports. Source bytes and Git HEAD remained unchanged. Different objectives can produce different baseline fold poses; pose parity across the ablation is not asserted.
| Drive, selected three inputs only | Scalar multiple-start controls detected | Orientation-proxy controls detected |
|---|---|---|
| 2011_09_26_0005 | 8/9 | 0/9 |
| 2011_09_28_0002 | 0/9 | 0/9 |
| Total | 8/18 | 0/18 |
These counts are control outcomes, not calibration accuracy, observed-axis coverage or native verdicts. Pitch +3 on 0002 remains undetected. The known roll +3 false control also remains rejected, but the scalar multiple-start method already rejected it. Losing every previously detected 0005 control does not constitute a useful confidence improvement. The scan-direction proxy is not recommended or promoted, and no threshold is relaxed after the result. No new SOTA conclusion follows from six deliberately selected spent inputs.
A separate replay of the spent 0005 baseline, with orientation/spatial flags
off and multiple starts on, matches the previous scalar report's full numerical
evidence excluding provenance within 1e-9. This checks that the revised
objective factory preserves that existing default path.
Receipts and next step¶
The batch, default confirmation and manifest bind schemas, all six probe outputs, the scalar replay and exact source/producer bytes, about 373 KB excluding the manifest. Images, point clouds, gradient tensors, NPZ payloads and bags remain local; no data was downloaded.
Tests cover gradient polarity, scale, weak/unknown orientation, rotated scan direction, signed-score behavior, margin support, unsupported inputs, frozen input/source bindings, the negative comparison and scalar-path parity. Broader orientation claims require genuine boundary geometry rather than assuming that scan direction is a boundary normal. Further development must also preserve turn sensitivity. Capture progress remains 2/55 spent drives; the remaining comparisons, exposure audit and independent pre-registration are still pending. Existing defaults and SOTA standings are unchanged.