Disjoint spatial prediction passes the controlled surface-quality screen¶
Frozen d72a83a1, evaluated 2026-10-30. See
protocol,
plan,
all54 paired rows,
schema, and
exact-plane validation.
The additional predictive model test passes the declared quality/support screen in 18/18 noisy frame conditions, versus10/18 for the fixed21px baseline. All nine clean conditions also pass, versus2/9 baseline. This supports the measurement-only surface construction on these controlled geometries; it does not yet establish correspondence accuracy, ICP recovery or real-data performance.
| Condition | Method | Median joint surface precision | Median domain retained | Frames passing all gates |
|---|---|---|---|---|
| Clean | Baseline21 | 74.65% | 33.34% | 2/9 |
| Clean | Predictive21 | 98.04% | 25.47% | 9/9 |
| Seed730 | Baseline21 | 80.94% | 30.65% | 5/9 |
| Seed730 | Predictive21 | 99.92% | 23.44% | 9/9 |
| Seed731 | Baseline21 | 81.69% | 30.81% | 5/9 |
| Seed731 | Predictive21 | 99.87% | 23.86% | 9/9 |
Noisy selected precision ranges99.52–99.99% (seed730) and99.63–99.98% (731). Retention ranges14.13–30.81% and14.46–31.07%. Every noisy frame also retains at least eight image cells and the declared normal-direction spectrum ratio. All 27 baseline rows exactly reproduce the archived21px result; paired measurement hashes match. The candidate changes only the selection mask, preserving the full 21px estimated depth/normal/confidence at retained pixels.
The exact-plane test accepts98.88% of10000 independent noise draws against the nominal99% predictive gate, passing the frozen98–99.7% sanity interval. It uses prediction covariance with shared coefficient uncertainty, not an independence assumption for residuals produced by the same fitted plane. The two checkerboard pixel sets have disjoint measurement noise under this synthetic model.
Limits and interpretation¶
The test fits221 pixels and predicts220 held-out pixels. It then selects locations whose original full441-pixel estimate is retained; final estimates reuse held-out pixels. This is model validation followed by refitting, not an independent final uncertainty audit. Checkerboard spatial samples are also not independent recordings. Real stereo errors may be correlated and may not have the assumed .5px variance.
Joint precision compares smoothed XYZ with exact analytic XYZ within .2m and unsigned normal error within15degrees. This is not proof that nearest-neighbor cross-modal pairings refer to the same physical surface, nor full six-DoF observability. Normal spectra are not covariance. Controlled layouts are spent development data, not fresh evaluation or an independent dataset family.
Observed angular coverage for selected noisy surfaces ranges94.40–98.57% for seed730 and93.65–96.91% for731. Clean coverage ranges91.26–97.00%. Therefore the noise-only angular bound must not be advertised as guaranteed95% coverage on every curved frame. Clean precision is95.52–99.07%, lower than noisy selected precision in some frames: selection distributions differ; this is not evidence that adding noise improves the underlying surface measurement.
Five focused tests, repository Ruff and strict documentation build pass. Tests cover exact-plane prediction, curvature rejection, missing data, unchanged selected estimates, empirical Gaussian predictive calibration and affine reconstruction. Validation checks schema,54-case completion, all frozen hashes, input identity and exact archived-baseline reproduction. No real-data read, calibration rerun, core/default change or SOTA claim occurred.
Next make the construction camera-parametric, verifying exact parity on these frozen controls. The pilot currently uses analytic-camera constants fx=fy240 and fx*baseline120; applying it verbatim to KITTI would be incorrect. Bind real camera intrinsics, stereo geometry and assumed noise explicitly in a new plan, account for re-rectification/resampling limitations, then assess the same18 spent real frames without fitting thresholds to reference normals. Only after that input quality check should a matched ICP comparison be frozen. Retain all rejected points/counts and failed frames; this synthetic pass is not adoption permission.