First-order depth uncertainty normalization fails recovery controls¶
Frozen at 2eb1336d before evaluation on 2026-10-30. See
protocol,
plan,
all 342 fits and
schema.
Reject this normalization candidate. It does not recover noisy perturbed poses and substantially regresses the clean controls. This is a test of one specific measured-depth, first-order uncertainty surrogate, not a rejection of range-dependent stereo uncertainty as a physical phenomenon.
| Noise condition | Start | Fixed joint pose+bias recovery | Normalized recovery |
|---|---|---|---|
| Clean | Perturbed | 12/36 | 1/36 |
| Clean | Reference | 6/18 | 0/18 |
| .5 px, seed 730 | Perturbed | 0/36 | 0/36 |
| .5 px, seed 730 | Reference | 0/18 | 3/18 |
| .5 px, seed 731 | Perturbed | 0/36 | 0/36 |
| .5 px, seed 731 | Reference | 0/18 | 5/18 |
On clean sphere cases there is one gain and 18 regressions, across perturbed and reference starts. All 63 clean fixed-baseline outputs reproduce the archived joint_pose_bias records exactly. Paired methods use byte-identical generated depth arrays, whose hashes are retained in every row. Thus the clean regression is not explained by a changed baseline, geometry or noise generation.
Noisy perturbed pose-only recovery is also 0/72 for both methods. Normalization improves noisy reference-start correction in eight cases, but reference starts are controls, not successful recovery from perturbation. Across noisy reference starts, median bias error changes from .5000 m for the fixed method to .05646 m (seed 730) and .05648 m (731) for the candidate, still beyond the .05 m gate in many cases. On clean perturbed starts, median bias error worsens .05729→.07060 m. Do not replace joint accuracy with favorable bias-only statistics.
Convergence and degeneracy¶
| Condition/start | Fixed convergence/support | Normalized convergence/support |
|---|---|---|
| Clean perturbed | 36/36 | 35/36 |
| Clean reference | 18/18 | 18/18 |
| Seed 730 perturbed | 25/36 | 0/36 |
| Seed 730 reference | 17/18 | 15/18 |
| Seed 731 perturbed | 25/36 | 2/36 |
| Seed 731 reference | 18/18 | 18/18 |
Both methods flag all nine clean parallel-plane controls as degenerate and reject them. Each method uses 8–800 actual optimizer calls; all caps hold, including the finite-difference evaluations. Total solve time including diagnostics was 136.36 s for fixed and 198.29 s for normalized; report serialization is excluded. This is not a controlled performance benchmark. All failed/nonconverged cases remain.
Interpretation and next decision¶
The candidate uses measured depth squared to approximate the variance induced by .5 px disparity noise, with a .1 m floor. It changes the effective metric robust loss and positive-occlusion scales as well as downweighting far measurements. Weights depend on projected/interpolated measurements. Therefore it is not an exact Gaussian disparity likelihood, and its behavior cannot be attributed to a single isolated change in covariance. The candidate assumes .5 px on clean data too; the clean controls explicitly test robustness to that mismatch.
The test confirms that simply normalizing metric residuals is insufficient under this protocol. It does not identify weighting, interpolation, asymmetric loss or nonlinear optimization as the sole cause, and it does not test correlated stereo errors, bad matches or a calibrated real sensor noise model. No real recordings were used to fit scales and no new real data were consumed.
Seven focused tests, repository Ruff and strict documentation build pass. Checks validate report/schema, all frozen source hashes, 342-case completion, paired depth hashes, budgets and exact 63-case clean baseline reproduction. Core/defaults and the real-data MI baseline remain unchanged; no adoption or SOTA claim.
Do not retune the floor or assumed sigma on these outcomes. Before another real transfer, the next distinct hypothesis is a direct disparity-domain measurement model, avoiding depth inversion's nonlinear error and measured-depth weighting. Specify it independently and freeze the same clean/noisy/occlusion/bias controls before running; retain separate numerical convergence and accuracy reporting. These results provide no evidence that the proposed next model already works.