Research map / Image generation / Image restoration
PolarFree: Polarization-based Reflection-Free Imaging
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.
Builds PolaRGB, a 6,500-pair real polarization/RGB reflection-removal dataset, and PolarFree, a two-stage diffusion-prior network with a phase loss that removes reflections from polarized images.
Problem
Single-image reflection removal is ill-posed; existing polarization-based methods rely on small, synthetic, or RGB-less datasets, limiting real-world generalization.
Approach
Uses a conditional DDPM to generate a reflection-free prior from polarization (AoLP, DoLP) and RGB inputs, supervised by a prior encoder trained on clean transmission, then feeds the prior into a reflection removal backbone; adds a phase-based frequency loss to reduce sensitivity to color mismatch.
Key results
- PolarFree reaches 22.44 PSNR, 0.8681 SSIM and 0.1325 LPIPS on PolaRGB, versus 20.67 PSNR for the best baseline (Kim et al.).
- Removing polarization cues drops PSNR from 22.44 to 21.56; removing the diffusion prior drops it to 20.21.
- Removing the phase loss reduces PSNR to 22.41 and worsens LPIPS (0.1402 vs 0.1325).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PolarFree instead of Kim et al. | PolaRGB reflection removal · PolaRGB test | PSNR: 20.67 -> 22.44 | Helped |
| Use PolarFree instead of Kim et al. | PolaRGB reflection removal · PolaRGB test | SSIM: 0.8399 -> 0.8681 | Helped |
| Use PolarFree instead of IBCLN | PolaRGB reflection removal · PolaRGB test | LPIPS: 0.2488 -> 0.1325 | Helped |
| Use PolarFree instead of Kim et al. | PolaRGB reflection removal · PolaRGB test | Q-align: 3.7148 -> 3.8867 | Helped |
| Add phase-based loss (ablation) | PolarFree · PolaRGB test | PSNR: 22.41 -> 22.44 | Helped |
| Add polarization inputs (ablation) | PolarFree · PolaRGB test | PSNR: 21.56 -> 22.44 | Helped |
| Add conditional diffusion prior (ablation) | PolarFree · PolaRGB test | PSNR: 20.21 -> 22.44 | Helped |
Surprising findings
- Baselines re-trained with polarization inputs (e.g. RDRNet) scored lower PSNR than older methods such as Kim et al., at 15.88 versus 20.67.
Limitations
- Real-world museum evaluation is qualitative only, with no ground truth.
- Test set is small (11 scenes, 188 pairs).
- Baselines were modified and retrained on PolaRGB, so comparisons depend on those adaptations.
Experimental setup
Datasets: PolaRGB
Benchmarks: PolaRGB
Metrics: PSNR SSIM LPIPS Q-Align
Baselines: Lei et al. Kim et al. IBCLN YTMT DSRNet RDRNet
Builds on: DDPM RDRNet Stokes parameters
Closest papers by meaning
- RGB-to-Polarization Estimation: A New Task and Benchmark Study (NeurIPS 2025 · similarity 0.89)
- Glossy Object Reconstruction with Cost-effective Polarized Acquisition (CVPR 2025 · similarity 0.88)
- PolarScale: A Physics-Grounded Benchmark for Radiometrically Consistent RGB-to-Stokes Estimation (NeurIPS 2026 · similarity 0.87)
- Flash-Split: 2D Reflection Removal with Flash Cues and Latent Diffusion Separation (CVPR 2025 · similarity 0.86)
- Polarization Guided Mask-Free Shadow Removal (AAAI 2025 · similarity 0.86)
- Spectral and Polarization Vision: Spectro-polarimetric Real-world Dataset (CVPR 2024 · similarity 0.85)
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.