Camera-parametric surface transfer improves fifteen-frame quality screen¶
Frozen 22453494, evaluated2026-10-30. See
protocol,
plan and camera geometry,
all36 paired rows and
schema.
The generic implementation reproduces every output array exactly for all27 analytic measurements. Real construction explicitly receives original camera intrinsics, calibration-derived fB and assumed .5px noise. Analytic constants are not used for the real frames. Invalid camera geometry is rejected.
| Across18 real frames | Baseline21 | Predictive21 |
|---|---|---|
| Frames passing all declared quality/support gates | 13/18 | 15/18 |
| Median retained domain | 26.81% | 24.77% |
| Retention range | 15.22–34.28% | 13.21–32.74% |
| Median target fraction within1m of LiDAR | 88.54% | 90.86% |
| Median normal agreement within15deg | 86.12% | 88.16% |
| Normal agreement range | 42.52–94.95% | 42.70–96.76% |
| Selected point count range | 66014–148728 | 57301–142016 |
Normal agreement is conditional on reference-pose nearest-LiDAR distance<=1m and a valid12-neighbor LiDAR normal. It is an evaluation proxy, not exact physical correspondence truth or calibration precision. Dense target pixels are scored; this is not the earlier voxel-target/source-FOV population. Spatial cells remain 13–16; valid normal pairs range45943–114861 for the candidate. Selection uses no reference or LiDAR normals, and predictive output is a subset of baseline points.
Every predictive frame¶
| Drive / frame | Retention | Within1m | Normal agreement | Pass |
|---|---|---|---|---|
| 2011_09_26_0005 / frame-38 | 13.21% | 80.57% | 42.70% | False |
| 2011_09_26_0005 / frame-76 | 21.56% | 91.81% | 84.21% | True |
| 2011_09_26_0005 / frame-114 | 25.52% | 68.62% | 76.28% | False |
| 2011_09_28_0002 / frame-93 | 26.33% | 87.84% | 88.50% | True |
| 2011_09_28_0002 / frame-187 | 30.65% | 88.92% | 91.39% | True |
| 2011_09_28_0002 / frame-281 | 31.35% | 87.82% | 93.26% | True |
| 2011_09_26_0009 / frame-111 | 26.01% | 91.15% | 76.43% | False |
| 2011_09_26_0009 / frame-223 | 29.54% | 87.08% | 85.73% | True |
| 2011_09_26_0009 / frame-334 | 24.68% | 92.41% | 81.68% | True |
| 2011_09_26_0011 / frame-58 | 31.26% | 90.98% | 93.47% | True |
| 2011_09_26_0011 / frame-116 | 24.86% | 95.51% | 96.00% | True |
| 2011_09_26_0011 / frame-174 | 32.74% | 93.83% | 96.76% | True |
| 2011_09_26_0013 / frame-35 | 24.21% | 90.75% | 87.82% | True |
| 2011_09_26_0013 / frame-71 | 23.79% | 88.33% | 93.50% | True |
| 2011_09_26_0013 / frame-107 | 20.68% | 95.89% | 81.07% | True |
| 2011_09_26_0015 / frame-74 | 23.46% | 88.69% | 92.41% | True |
| 2011_09_26_0015 / frame-148 | 21.24% | 92.82% | 84.94% | True |
| 2011_09_26_0015 / frame-222 | 20.19% | 91.77% | 91.46% | True |
All-frame continuation fails. The remaining frames are0005/frame38, 0005/frame114 and0009/frame111; normal agreement is42.70%,76.28%,76.43% respectively. Frame114 also has only68.62% within1m. No frame is dropped, and the80% screen is not relaxed. No ICP rerun is performed as specified by the prior protocol.
The surrogate disparity is fB/z on restored original-camera depths. Rectification, nearest-neighbor restoration and stereo matching can violate the independent Gaussian assumption. Consequently the predictive chi-square threshold is a fixed transfer hypothesis, not calibrated real99% confidence. LiDAR normals can also be wrong at boundaries or with motion; these results do not uniquely attribute the three failures to stereo, timing or source geometry. Surface similarity alone does not demonstrate observable or recoverable six-DoF extrinsics.
Five focused tests, repository Ruff and strict documentation build pass. Checks verify schema, all36 rows, every source/input/calibration hash and27-case bitwise analytic parity for every estimator output array. Tests cover changed focal/stereo scale, analytic plane depth/normal, invalid geometry and the prior predictive gate. No core/default change, newly acquired data, calibration accuracy or SOTA claim.
Next inspect the spatial distribution of failures in all18 frames, emphasizing the three failed ones, using fixed reference-only distance/normal residual maps and measurement-only support masks. Determine whether errors concentrate at dynamic/ depth boundaries or large regions before changing the measurement model. Preserve the failed frames and avoid making truth-dependent exclusion masks. A matched registration experiment remains pending; the synthetic pass and15/18 quality screen do not establish a real-data calibration improvement.