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

Beibei Lin; Tingting Chen; Xin Zhang; Wenhao Zhao; Dongjun Li; Zifeng Yuan

cs.CVOct 6, 2026

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

Authors: Beibei Lin, Tingting Chen, Xin Zhang, Wenhao Zhao, Dongjun Li, Zifeng Yuan

Organizations: National University of Singapore · University of Michigan, Ann Arbor

Abstract

Polarization imaging provides physical cues beyond intensity imaging but typically requires specialized hardware. Recent methods infer polarization from RGB-like inputs, yet predict only normalized Stokes components or relative descriptors, from which the radiometric scale needed for full Stokes reconstruction has been divided out. We introduce PolarScale, a benchmark that makes this scale an explicit prediction and evaluation target. Built on existing trichromatic full-Stokes measurements, PolarScale takes the per-scene normalized total-intensity image $s_0$ (a scene-referred linear image, not a consumer sRGB photograph) and asks models to predict normalized Stokes components, AoLP/DoLP/DoCP, and a per-scene scale. Because the scale is divided out of the input, it is not physically identifiable; PolarScale therefore evaluates dataset-conditioned semantic scale estimation against a constant-scale control, together with angular, self-consistency, and physical-bound metrics. Across seven restoration-based and generative backbones and three prediction strategies, the strongest restoration models estimate the scale with 3.6-4.3% mean relative error versus 5.7% for the constant control and violate physical bounds on fewer than 0.25% of pixels, whereas two generative baselines collapse to a near-zero scale; explicit descriptor supervision improves descriptor accuracy (23.66 vs. 18.88 dB PSNR for MAE). Predicted full-Stokes representations improve diffuse/specular separation, material segmentation, and glare classification, although in diffuse/specular separation the learned scale performs only on par with the constant control.

Figures & tables

Figure 1: Overview of PolarScale . Prior RGB-based methods [ 22 , 21 ] predict normalized Stokes components or relative polarimetric descriptors, from which the radiometric scale has been divided out. PolarScale additionally predicts the per-scene radiometric scale $\sigma$ and reconstructs the full Stokes components $\mathrm{S}_{n}=\sigma\,\mathrm{s}_{n}$ , making the scale an explicit evaluation target. The input is the normalized total-intensity image $\mathrm{s}_{0}=\mathrm{S}_{0}/\sigma$ (Section 3.2 ), not a consumer sRGB photograph.

Figure 2: Overview of the three prediction strategies evaluated in our benchmark. (a) Direct Stokes Estimation predicts $\sigma$ and normalized Stokes components, then derives AoLP, DoLP, and DoCP. (b) Joint Estimation predicts all polarimetric parameters with a unified decoder. (c) Decoupled Multi-decoder separately estimates scale-dependent components and scale-independent descriptors. The input in all three strategies is the normalized total-intensity image $\mathrm{s}_{0}$ .

Figure 3: Reconstructed full Stokes components for test scene 1043 (input: normalized $\mathrm{s}_{0}$ ). The sky carries strong $\mathrm{S}_{2}$ and $\mathrm{S}_{3}$ signals: MAE [ 33 ] recovers their extent and sign, Uformer [ 32 ] only faint edges, and DiT [ 35 ] and I2ITurbo [ 37 ] predict nearly flat maps.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Meaning
$\mathrm{S}_{0},\mathrm{S}_{1},\mathrm{S}_{2},\mathrm{S}_{3}$	Full (absolute) Stokes components, trichromatic ( $H\times W\times 3$ each)
$\sigma$	Per-scene radiometric scale, $\sigma=\max\mathrm{S}_{0}$
$\hat{\sigma}$	Predicted per-scene scale
$\mathrm{s}_{k}=\mathrm{S}_{k}/\sigma$	Normalized Stokes components, $k\in\{0,1,2,3\}$
$\mathrm{s}_{0}$	Network input: normalized total-intensity image
$\hat{\mathrm{s}}_{n}$	Predicted normalized Stokes components, $n\in\{1,2,3\}$

Appendix

Table 6: Notation used in this paper.

Figure 4: Reconstructed full Stokes components for test scene 1050, an urban plaza with high-rise buildings in which the sky region carries a strong negative $\mathrm{S}_{2}$ signal. MAE recovers $\mathrm{S}_{2}$ and $\mathrm{S}_{3}$ in this region but predicts the wrong sign for $\mathrm{S}_{1}$ , while the other models predict nearly flat maps.

Figure 5: Reconstructed full Stokes components for test scene 1003.

Figure 6: Reconstructed full Stokes components for test scene 1035.

Figure 7: Reconstructed full Stokes components for test scene 1038.

Figure 8: Reconstructed full Stokes components for test scene 1062.

Figure 9: Reconstructed full Stokes components for test scene 1083.

Figure 10: Reconstructed full Stokes components for test scene 1126.

Figure 11: Predicted AoLP, DoLP, and DoCP for test scene 1158, a street intersection with crosswalk stripes, vehicles, and building facades. The models capture the global layout but miss abrupt polarization transitions; DiT produces noisy AoLP and I2ITurbo saturates DoCP, while MAE is closest to the ground truth.

Figure 12: Predicted AoLP, DoLP, and DoCP for test scene 1196, a high-rise facade whose repetitive window pattern produces fine, periodic polarization structure.

Figure 13: Predicted AoLP, DoLP, and DoCP for test scene 1036.

Figure 14: Predicted AoLP, DoLP, and DoCP for test scene 1040.

Figure 15: Predicted AoLP, DoLP, and DoCP for test scene 1047.

Figure 16: Predicted AoLP, DoLP, and DoCP for test scene 1052.

Figure 17: Predicted AoLP, DoLP, and DoCP for test scene 1194.

Explore similar work

Stokes-Informed Diffusion for Robust Linear Polarization Estimation
Jul 23, 2026Yidong Luo, Chenggong Li, Yuchao Feng +3Infrared-Visible Image Fusion Diffusion Models
Zhejiang University, Hangzhou, China · School of Automation, Central South University, Changsha, China · College of Computer Science and Technology, Zhejiang University of Technology, Hangzhou, China +2
DensePol: Dense-Angle Polarization Dataset for Learning-Based Polarimetric Vision
Sep 8, 2026Param Sangani, Ahmad Moori, Erik Blasch +2Infrared-Visible Image Fusion Vision Datasets
Department of Computer Science, Saint Louis University, St. Louis, MO, USA · MOVEJ Analytics, Fairborn, OH, USA · U.S. Naval Research Laboratory, Washington, DC, USA
PolarAPP: Beyond Polarization Demosaicking for Polarimetric Applications
Mar 24, 2026Yidong Luo, Chenggong Li, Yunfeng Song +4Infrared-Visible Image Fusion Computer Vision
Zhejiang University, Hangzhou, China · School of Automation, Central South University, Changsha, China · School of Engineering, Westlake University, Hangzhou, China +1

Jul 23, 2026, cs.CVJ/K move · Enter open · S save

Stokes-Informed Diffusion for Robust Linear Polarization Estimation

Yidong Luo, Chenggong Li, Yuchao Feng, Boxin Shi+2

Zhejiang University, Hangzhou, China · School of Automation, Central South University, Changsha, China · College of Computer Science and Technology, Zhejiang University of Technology, Hangzhou, China+2

Infrared-Visible Image Fusion Diffusion Models Reynolds-Averaged Navier-Stokes

Open paper AskPDF arXiv

Polarization cues benefit applications such as material detection and de-reflection, yet acquiring them typically requires dedicated hardware. This motivates us to estimate the linear polarization from a single RGB image. However, the task is inherently ill-posed, with the Angle of Polarization (AoP) becoming particularly unstable in weak polarization regions, where the polarimetric signal is overwhelmed by noise, leading to erratic angle estimates. To address these limitations, we propose GenPolar, a Stokes-informed diffusion framework grounded in the Mueller formalism from an intensity observation. Specifically, GenPolar predicts channel-wise linear Stokes components (S1,S2) from intensity S0, from which degree of linear polarization (DoLP) and AoP are analytically derived; AoP is further supervised with an observability-aware loss. In addition, to enable efficient and high-fidelity inference, we adopt a two-stage training strategy. Firstly, a multi-step conditional diffusion model is trained with a physics-based loss. Subsequently, we distill it into a one-step generator, which further supports stable Low-Rank Adaptation (LoRA) of the VAE encoder to mitigate domain-specific autoencoding bias. Extensive experiments across rotating-polarizer, division-of-focal-plane, and hybrid datasets demonstrate that GenPolar achieves state-of-the-art performance in both DoLP fidelity and AoP stability. Crucially, these improvements translate to significant and consistent gains in downstream applications, including material detection and de-reflection.

Abstract

Explore similar work

Stokes-Informed Diffusion for Robust Linear Polarization Estimation

DensePol: Dense-Angle Polarization Dataset for Learning-Based Polarimetric Vision

PolarAPP: Beyond Polarization Demosaicking for Polarimetric Applications