Research map / Image generation / Image restoration
Polarization Guided Mask-Free Shadow Removal
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.
Proposes Pol-ShaRe, a polarization-guided, mask-free two-stage shadow removal network that uses intensity and DoP-based priors and a modulated total intensity intermediate.
Problem
Image shadow removal typically requires an externally obtained shadow mask, which adds workload and degrades performance near boundaries; mask-free methods suffer from artifacts due to ill-posedness.
Approach
Uses polarization (degree of polarization) together with intensity-domain priors to indicate per-pixel shadow confidence without a mask, decouples restoration into two stages (modulated total intensity m then image I), and uses DDPF and TGD modules to suppress artifacts.
Key results
- Pol-ShaRe achieves 32.24 PSNR, 0.981 SSIM and 2.858 RMSE on all pixels of synthetic data, outperforming TBRNet (29.13 PSNR) and ShadowFormer (31.16 PSNR) which were given ground-truth masks.
- Ablation: removing polarization drops PSNR from 32.24 to 27.83; single-stage variant drops to 29.06.
- Qualitatively produces less false color on real captured images than mask-based and mask-free baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add polarization information (ablation) | Pol-ShaRe · Synthetic data, all pixels | PSNR: 27.83 -> 32.24 | Helped |
| explicitly extract the priors (ablation) | Pol-ShaRe · Synthetic data, all pixels | PSNR: 28.75 -> 32.24 | Helped |
| use Ppol instead of a shadow mask (ablation) | Pol-ShaRe · Synthetic data, all pixels | PSNR: 30.06 -> 32.24 | Helped |
| use the two-stage pipeline instead of a single stage (ablation) | Pol-ShaRe · Synthetic data, all pixels | PSNR: 29.06 -> 32.24 | Helped |
| use DDPF instead of convolution layers (ablation) | Pol-ShaRe · Synthetic data, all pixels | PSNR: 31.48 -> 32.24 | Helped |
| use TGD instead of convolution layers (ablation) | Pol-ShaRe · Synthetic data, all pixels | PSNR: 30.92 -> 32.24 | Helped |
| use polarization-guided Pol-ShaRe rather than ShadowFormer | Synthetic data · Synthetic data, all pixels | PSNR: 31.16 -> 32.24 | Helped |
Limitations
- Polarization prior fails in indoor scenes with artificial light or adverse weather such as rain, haze, or snow.
- Requires multiple polarized shots, so misalignment can degrade performance in dynamic scenes.
Experimental setup
Datasets: own synthetic polarization shadow dataset
Benchmarks: own synthetic dataset real captured images
Metrics: PSNR SSIM RMSE in LAB space
Baselines: Inpaint4Shadow ShadowDiffusion ShadowFormer DMTN TBRNet
Builds on: Malus's law polarization image formation attention gate ShadowFormer Ono et al. 2022 color constancy weighting
Closest papers by meaning
- Diff-Shadow: Global-guided Diffusion Model for Shadow Removal (AAAI 2025 · similarity 0.88)
- Recasting Regional Lighting for Shadow Removal (AAAI 2024 · similarity 0.88)
- SoftShadow: Leveraging Soft Masks for Penumbra-Aware Shadow Removal (CVPR 2025 · similarity 0.88)
- When Shadow Removal Meets Intrinsic Image Decomposition: A Joint Learning Framework Using Unpaired Data (AAAI 2025 · similarity 0.88)
- Detail-Preserving Latent Diffusion for Stable Shadow Removal (CVPR 2025 · similarity 0.87)
- DeS3: Adaptive Attention-Driven Self and Soft Shadow Removal Using ViT Similarity (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.