Accurate radiometric scale estimation from normalized intensity images

Improve the accuracy of per-scene radiometric scale estimation from normalized total-intensity images so that predicted full-Stokes representations can consistently outperform a constant dataset-prior scale in downstream tasks.

Background

PolarScale predicts normalized Stokes components and a per-scene radiometric scale from an input image in which the true scale has been divided out. Consequently, the scale cannot be recovered through physical inversion and must instead be inferred from scene content and dataset-conditioned semantic cues. The benchmark compares learned scale estimates with a constant-scale control representing the training-set prior.

Although the strongest restoration models achieve lower scale error than the constant control, the learned scale does not consistently improve downstream performance: in diffuse/specular separation it performs on par with the constant scale, while the ground-truth scale yields a clear improvement. The authors therefore identify more accurate scale estimation as an unresolved problem for future physically grounded RGB-to-Stokes models.

References

Our claim is therefore that predicted full-Stokes representations benefit the evaluated tasks, with the absolute scale as one contributing component whose accuracy remains the main open problem.

— PolarScale: A Physics-Grounded Benchmark for Radiometrically Consistent RGB-to-Stokes Estimation  (2610.08346 - Lin et al., 6 Oct 2026) in Section 5.2, subsection “Downstream validation,” paragraph “Results”; also identified in Section 1, paragraph preceding “Our main contributions are summarized as follows”