High-Resolution Single-Shot Polarimetric Imaging Made Easy

Shuangfan Zhou; Chu Zhou; Heng Guo; Youwei Lyu; Boxin Shi; Zhanyu Ma; Imari Sato

cs.CVApr 7, 2026

High-Resolution Single-Shot Polarimetric Imaging Made Easy

Authors: Shuangfan Zhou, Chu Zhou, Heng Guo, Youwei Lyu, Boxin Shi, Zhanyu Ma, Imari Sato

Organizations: School of Artificial Intelligence, Beijing University of Posts and Telecommunications, China · National Institute of Informatics, Japan · State Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University, China · National Key Laboratory of General AI, School of Intelligence Science and Technology, Peking University, China

Abstract

Polarization-based vision has gained increasing attention for providing richer physical cues beyond RGB images. While achieving single-shot capture is highly desirable for practical applications, existing Division-of-Focal-Plane (DoFP) sensors inherently suffer from reduced spatial resolution and artifacts due to their spatial multiplexing mechanism. To overcome these limitations without sacrificing the snapshot capability, we propose EasyPolar, a multi-view polarimetric imaging framework. Our system is grounded in the physical insight that three independent intensity measurements are sufficient to fully characterize linear polarization. Guided by this, we design a triple-camera setup consisting of three synchronized RGB cameras that capture one unpolarized view and two polarized views with distinct orientations. Building upon this hardware design, we further propose a confidence-guided polarization reconstruction network to address the potential misalignment in multi-view fusion. The network performs multi-modal feature fusion under a confidence-aware physical guidance mechanism, which effectively suppresses warping-induced artifacts and enforces explicit geometric constraints on the solution space. Experimental results demonstrate that our method achieves high-quality results and benefits various downstream tasks.

Figures & tables

Figure 1: Comparison of representative polarimetric imaging systems. (a) DoT system : a classical Division-of-Time setup that provides high-quality polarization measurements via sequential acquisition, but fails in dynamic scenes due to motion-induced misalignment. (b) DoFP camera : enables snapshot capture through pixel-level multiplexing, at the cost of reduced spatial resolution and demosaicing artifacts. (c) PolarAnything : a generative approach ( Zhang et al., 2025 ) that predicts polarization from a single RGB image, which may lead to physically inconsistent results. (d) EasyPolar (ours) : achieves high-fidelity snapshot imaging by preserving a full-resolution, high-SNR unpolarized reference, avoiding the fundamental resolution and noise trade-offs of DoFP sensors while maintaining physical consistency.

Figure 2: Overview of the proposed physics-informed reconstruction network. Given three input images with distinct viewpoints and polarization states, we first perform viewpoint alignment and surface normal estimation using a pre-trained disparity network. The framework then employs a multi-modal feature fusion module to extract and aggregate features from polarization, texture, and geometry. Subsequently, a confidence-gated reconstruction module uses pseudo-polarization priors to refine the fused features, while adaptively controlling the influence of these priors via an estimated confidence map $\mathbf{C}$ . By combining these components, the network leverages physical constraints and confidence guidance to produce high-fidelity AoP and DoP estimates.

Figure 3: Our real-world data acquisition setup. The prototype comprises three synchronized RGB cameras mounted with a narrow horizontal baseline. The left and right cameras are equipped with fixed linear polarizers oriented at $0^{\circ}$ and $45^{\circ}$ , respectively. The middle camera temporarily fitted with a motorized rotatable polarizer only during data collection to enable time-multiplexed measurements for high-precision ground-truth computation. During inference, this polarizer is removed, and the central view serves as an unpolarized reference for single-shot capture.

	Method	$I_{0}$	$I_{45}$	$S_{1}$	$S_{2}$	$\theta$	$\rho$
	Method	PSNR $\uparrow$ /SSIM $\uparrow$	PSNR $\uparrow$ /SSIM $\uparrow$	PSNR $\uparrow$ /SSIM $\uparrow$	PSNR $\uparrow$ /SSIM $\uparrow$	MAE $\downarrow$	PSNR $\uparrow$ /SSIM $\uparrow$
Syn Data	DoFP camera + PIDSR ( Zhou et al., 2025d )	41.39 / 0.9788	41.00 / 0.9787	50.84 / 0.9953	52.20 / 0.9963	5.97	32.80 / 0.9328
	Polaranything ( Zhang et al., 2025 )	27.86 / 0.8023	28.63 / 0.8263	27.75 / 0.6580	28.83 / 0.6935	45.32	17.82 / 0.1707
	Ours	41.42 / 0.9965	42.15 / 0.9972	51.02 / 0.9955	53.16 / 0.9965	3.28	33.15 / 0.9483
Real Data	DoFP camera + PIDSR ( Zhou et al., 2025d )	35.22 / 0.9148	35.04 / 0.9103	40.30 / 0.9687	44.42 / 0.9779	11.97	26.27 / 0.7403
	Polaranything ( Zhang et al., 2025 )	34.61 / 0.9034	36.07 / 0.9171	34.89 / 0.8807	35.88 / 0.8910	44.76	19.71 / 0.2314

Table 1: Quantitative comparisons on synthetic data and real data. The comparisons involve EasyPolar, hardware-based method (DoFP + PIDSR ( Zhou et al., 2025d ) ) and generative baseline (Polaranything ( Zhang et al., 2025 ) )

Figure 4: Qualitative comparison on synthetic data between EasyPolar, PolarAnything ( Zhang et al., 2025 ) , and the DoFP camera results processed by a state-of-the-art polarization demosaicing and SR method (PIDSR ( Zhou et al., 2025d ) ).

Figure 5: Qualitative comparison on real-world data between EasyPolar, PolarAnything ( Zhang et al., 2025 ) , and the DoFP camera results processed by a state-of-the-art polarization demosaicing and SR method (PIDSR ( Zhou et al., 2025d ) ).

Method	$I_{0}$	$I_{45}$	$S_{1}$	$S_{2}$	$\theta$	$\rho$
Method	PSNR $\uparrow$ /SSIM $\uparrow$	PSNR $\uparrow$ /SSIM $\uparrow$	PSNR $\uparrow$ /SSIM $\uparrow$	PSNR $\uparrow$ /SSIM $\uparrow$	MAE $\downarrow$	PSNR $\uparrow$ /SSIM $\uparrow$
(1) intensity regression	24.97 / 0.9250	26.57 / 0.9415	24.96 / 0.8989	26.57 / 0.8943	29.09	12.22 / 0.2217
(2) w/o AoP trigonometric encoding	40.05 / 0.9658	43.14 / 0.9824	39.84 / 0.9651	43.25 / 0.9782	11.24	26.01 / 0.7402
(3) w/o geometry encoder	40.26 / 0.9684	43.57 / 0.9851	40.08 / 0.9682	43.81 / 0.9795	10.87	25.85 / 0.7357
(4) w/o confidence gated	39.80 / 0.9648	43.06 / 0.9887	39.62 / 0.9665	43.48 / 0.9794	12.10	25.43 / 0.7294
(5) w/ $L_{1}$ loss	40.10 / 0.9693	43.43 / 0.9878	39.75 / 0.9672	43.62 / 0.9813	10.62	25.69 / 0.7307

Table 2: Quantitative evaluation results of ablation study.

Figure 6: Qualitative results of the ablation study. The results demonstrate that our EasyPolar yields the most accurate polarization details and geometric structures compared to the ablated versions.

Figure 7: Robustness to camera baseline variations. We evaluate our method under three different baseline distances. With the proposed view encoding, our full framework achieves accurate AoP and DoP estimates with minimal variance across baselines, whereas the version without view encoding fails to handle larger disparities, leading to noticeable artifacts.

Figure 8: Qualitative comparison on polarization-enhanced downstream tasks. ( Top ) Glass segmentation: Using our estimated polarization images, a standard glass segmentation method produces cleaner masks that align more closely with the reference (derived from GT polarimetric images) compared to other baselines. ( Bottom ) Shape from polarization: The mean angular error (MAE) of surface normals, estimated using the method of Lyu et al. (2024) , is annotated below each result. The MAE of EasyPolar is the most consistent with that of the GT polarization, demonstrating our method’s superior fidelity in preserving physical polarization cues.

Explore similar work

PolarVSR: A Unified Framework and Benchmark for Continuous Space-Time Polarization Video Reconstruction
May 11, 2026Chenggong Li, Yidong Luo, Junchao Zhang +2Image Reconstruction Polarimetric Imaging
School of Automation, Central South University, Changsha, China · Hunan Provincial Key Laboratory of Optic-Electronic Intelligent Measurement and Control, Changsha, China · Zhejiang University, Hangzhou, China +3
Architectural Unification for Polarimetric Imaging Across Multiple Degradations
Mar 6, 2026Chu Zhou, Yufei Han, Junda Liao +6Polarimetric Imaging Image Restoration
Digital Content and Media Sciences Research Division, National Institute of Informatics, Tokyo 101-8430, Japan. · Pattern Recognition and Intelligent System Laboratory, School of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing 100876, China · State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University, Beijing 100080, China, and also with the National Engineering Research Center of Visual Technology, School of Computer Science, Peking University, Beijing 100080, China.
Guided Lensless Polarization Imaging
Mar 28, 2026Noa Kraicer, Erez Yosef, Raja GiryesPolarimetric Imaging Computational Imaging
Tel Aviv University School of Electrical Engineering, Faculty of Engineering

May 11, 2026J/K move · Enter open · S save

PolarVSR: A Unified Framework and Benchmark for Continuous Space-Time Polarization Video Reconstruction

Chenggong Li, Yidong Luo, Junchao Zhang, Boxin Shi+1

School of Automation, Central South University, Changsha, China · Hunan Provincial Key Laboratory of Optic-Electronic Intelligent Measurement and Control, Changsha, China · Zhejiang University, Hangzhou, China+3

Image Reconstruction Polarimetric Imaging

Open paper AskPDF arXiv

Polarimetric imaging captures surface polarization characteristics, such as the Degree of Linear Polarization (DoLP) and the Angle of Polarization (AoP). In mainstream Division of-Focal-Plane (DoFP) color polarization imaging, recovering polarization parameters from captured mosaic arrays remains a challenging inverse problem. Existing DoFP cameras also face hardware bottlenecks and often cannot support high-frame-rate acquisition, limiting polarimetric imaging in dynamic video tasks. These limitations motivate joint spatial and temporal enhancement. To this end, we propose the first space-time polarization video reconstruction architecture. The method jointly models polarization directions in space and time and uses a polarization-aware implicit neural representation for continuous, high-fidelity upsampling. By analyzing temporal variations in polarization parameters, we further introduce a flow-guided polarization variation loss to supervise polarization dynamics. We also establish the first large-scale color DoFP polarization video benchmark to support this research direction. Extensive experiments on this benchmark demonstrate the effectiveness of the method.

Abstract

Explore similar work

PolarVSR: A Unified Framework and Benchmark for Continuous Space-Time Polarization Video Reconstruction

Architectural Unification for Polarimetric Imaging Across Multiple Degradations

Guided Lensless Polarization Imaging