Direct metric-depth fitting results¶
Protocol frozen at d6651dd8. All 15 MI and 15 metric
fits completed. Every recomputed MI result exactly matches the original analytic
control. All metric fits satisfy numerical convergence and support retention.
| Condition | Seed | MI deg / m | MI pass | Metric deg / m | Metric pass | Residual calls | Holdout pass |
|---|---|---|---|---|---|---|---|
| exact_z_visible | +1 | 1.0459 / 0.2160 | False | 0.1415 / 0.0363 | True | 127 | True |
| exact_z_visible | -1 | 0.8972 / 0.1700 | False | 0.1433 / 0.0366 | True | 111 | True |
| exact_z_visible | +0 | 0.0349 / 0.0063 | True | 0.1414 / 0.0363 | True | 142 | True |
| exact_z_all | +1 | 1.0717 / 0.2349 | False | 0.7519 / 0.1054 | False | 121 | False |
| exact_z_all | -1 | 1.3748 / 0.2359 | False | 0.7518 / 0.1055 | False | 145 | False |
| exact_z_all | +0 | 0.0271 / 0.0000 | True | 0.7523 / 0.1054 | False | 142 | False |
| exact_range_visible | +1 | 0.4725 / 0.0718 | True | 0.1433 / 0.0366 | True | 139 | True |
| exact_range_visible | -1 | 0.8184 / 0.2767 | False | 0.1433 / 0.0366 | True | 134 | True |
| exact_range_visible | +0 | 0.2500 / 0.0000 | True | 0.1414 / 0.0363 | True | 142 | True |
| biased_z_visible | +1 | 1.0459 / 0.2160 | False | 0.3831 / 0.2222 | False | 202 | False |
| biased_z_visible | -1 | 0.8972 / 0.1700 | False | 0.3843 / 0.2224 | False | 144 | False |
| biased_z_visible | +0 | 0.0349 / 0.0063 | True | 0.3843 / 0.2224 | False | 158 | False |
| holes_z_visible | +1 | 1.1449 / 0.2278 | False | 0.1091 / 0.0329 | True | 132 | True |
| holes_z_visible | -1 | 1.2791 / 0.2236 | False | 0.1099 / 0.0331 | True | 157 | True |
| holes_z_visible | +0 | 0.0884 / 0.0153 | True | 0.1025 / 0.0318 | True | 127 | True |
Interpretation¶
Perturbed recovery improves MI 1/10 → metric 6/10, versus 5/10 for the separately recorded continuation experiment. Exact visible optical z, correctly converted radial depth and missing-rectangle conditions each recover 2/2. Both signed initializations reach similar solutions, unlike continuation's positive-only successes. Visible exact-depth errors remain about 0.14 degrees / 3.6 cm: analytic point intersections and pixel-interpolated depth are not identical near silhouettes, and no claim of an unbiased/global optimum is established.
The all-return camera-occlusion condition recovers 0/2, with approximately 0.752-degree rotation error. The +0.25 m depth bias condition recovers 0/2, with approximately 0.222 m translation error. Reference-start controls fall from MI's 5/5 to 3/5: those same two conditions drift outside the pose gates even from truth. All six of these failed metric outputs nevertheless converge numerically and retain support. Robust loss alone does not solve visibility mismatch or measurement bias, and numerical success remains unsafe as an accuracy criterion.
The frozen held-out consistency checks pass all nine accurate metric outputs and reject all six inaccurate ones in this run. This descriptive agreement does not repair the validator's previously demonstrated false accepts on other candidates. It must not be promoted or treated as independent ground truth. The same scene and camera-depth generation underpin fitting and diagnostic samples.
Actual residual evaluations (including numerical Jacobians) are 111–202, below the hard 800 ceiling. No retries, timeouts, budget failures or support failures. Total measured fitting time for 15 cases: MI 6.39 s, metric 4.12 s on this CPU. Different objectives, interpolation, optimizers and effective caps mean this is not an equal-work solver benchmark or an isolated objective ablation. The metric adapter also computes support counts; held-out checks are outside its timed solve block. Scene/conditions are dependent development controls, not independent real-world successes.
Decision¶
Retain the metric estimator as experimental. No production promotion or SOTA claim: the declared occlusion and reference-drift controls still fail. Preserve both successes and the six converged inaccurate results. Do not soften the 0.20 m gate for the 0.222 m bias outcome or prune points using reference visibility.
The next investigation should explicitly separate unobserved/occluded geometry from camera-depth bias, with predeclared controls on additional analytic scenes. A physically motivated visibility model must be tested without oracle extrinsics; a bias parameter must be checked for confounding with extrinsic translation. Do not add both mechanisms at once or tune them against held-out acceptance labels. First establish whether an occlusion-aware metric objective preserves visible-only recovery while avoiding the all-return truth-start drift, then address bias.
Full results and schema retain every status, pose, support count, residual-call count, runtime and held-out diagnostic. Seven focused tests, repository Ruff, strict documentation build, report/source-hash validation and exact baseline reproduction pass. Core code/defaults are unchanged.