Research map / Image generation / Image restoration

PolarScale: A Physics-Grounded Benchmark for Radiometrically Consistent RGB-to-Stokes Estimation

NeurIPS 2026 · Image generation › Image restoration · approach: Image harmonization and color transformation

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces PolarScale, a benchmark that makes the per-scene radiometric scale an explicit prediction and evaluation target for RGB-to-Stokes estimation, evaluated on seven restoration and generative backbones under three prediction strategies.

Problem

Existing RGB-based polarization inference predicts only normalized Stokes components or relative descriptors, with the radiometric scale divided out, so absolute Stokes vectors cannot be evaluated or recovered.

Approach

Treat the per-scene scale sigma = max S0 as an explicit target, reconstruct full Stokes as S_n = sigma * s_n, and evaluate scale estimation against a constant-scale control, alongside wrap-aware AoLP error, Stokes self-consistency, and physical-bound violation metrics.

Key results

• MAE achieves 3.6% mean relative scale error vs 5.7% for the constant-median control, with 0.18% bound violations.
• Explicit descriptor supervision with decoupled decoding raises MAE descriptor PSNR from 18.88 dB (direct) to 23.66 dB (decoupled).
• Predicted absolute Stokes raises glare-level accuracy from 48.5% to 63.0%, while the learned scale only matches the constant control in diffuse/specular separation (0.118/0.445 vs 0.118/0.448 AbsRel).

Surprising findings

• Full-Stokes PSNR barely depends on whether the scale is learned; a constant scale matches or exceeds the learned scale for well-behaved models.
• Two generative baselines (WDiff, DiT) collapse to a near-zero scale, with non-positive predictions on 11 and 97 of 200 test scenes.

Limitations

• Scale is not physically identifiable, so predictions are dataset-conditioned semantic estimates rather than calibrated radiometry.
• Single source dataset captured with one polarimetric camera and a narrow scale distribution; a constant predictor already reaches 5.7% error.
• Single training run per configuration; downstream probes are small-scale and some lack scale controls.

Experimental setup

Datasets: Spectral-Polarization dataset MCubeS

Benchmarks: PolarScale MCubeS

Metrics: PSNR SSIM LPIPS relative scale error log-scale error AoLP angular error self-consistency bound violation rate AbsRel mIoU

Baselines: Restormer Uformer MAE WDiff DiT RealFill I2ITurbo constant-scale control

Builds on: Spectral-Polarization dataset Restormer Uformer MAE RealFill I2ITurbo

Closest papers by meaning

• RGB-to-Polarization Estimation: A New Task and Benchmark Study (NeurIPS 2025 · similarity 0.90)
• Spectral and Polarization Vision: Spectro-polarimetric Real-world Dataset (CVPR 2024 · similarity 0.87)
• PolarFree: Polarization-based Reflection-Free Imaging (CVPR 2025 · similarity 0.87)
• Glossy Object Reconstruction with Cost-effective Polarized Acquisition (CVPR 2025 · similarity 0.86)
• NeISF: Neural Incident Stokes Field for Geometry and Material Estimation (CVPR 2024 · similarity 0.84)
• High-Fidelity Polarimetric Implicit 3D Reconstruction with View-Dependent Physical Representation (AAAI 2025 · similarity 0.84)