Research map / Image generation / Image restoration
Revisiting Single Image Reflection Removal In the Wild
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 a video-based pipeline for collecting aligned real reflection pairs, the RRW dataset (14,952 pairs), and a location-aware cascaded network (RDNet + RRNet) guided by the Maximum Reflection Filter (MaxRF) for single-image reflection removal.
Problem
Single-image reflection removal in real-world conditions is limited by misaligned or artifact-prone real training pairs collected with prior pipelines, and by indirect or implicit estimation of local reflection locations.
Approach
Block reflected light with black cloth in video capture to get aligned reflection-free and reflection images, then define explicit local reflection locations as pixels where Sobel gradient of the reflection image exceeds that of the transmission image (MaxRF), and use them to guide a two-stage removal network.
Key results
- Achieves average PSNR 25.40 dB across Nature, Real and SIR2, the best among 11 compared SIRR methods.
- Reaches 25.96 dB on Nature and 25.45 dB on SIR2; on Real it reaches 23.82 dB.
- Adding RDNet and RRW extra data improves ERRNet on Nature from 22.57 to 24.15 dB PSNR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add RRW extra training data to RDNet + RRNet (ablation) | RDNet + RRNet · Nature | PSNR: 25.37 -> 25.96 | Helped |
| Add RRW extra training data to RDNet + RRNet (ablation) | RDNet + RRNet · Real | PSNR: 22.65 -> 23.82 | Helped |
| Add RRW extra training data to RDNet + RRNet (ablation) | RDNet + RRNet · SIR2 | PSNR: 24.93 -> 25.45 | Helped |
| Add RDNet reflection-location guidance to RRNet + RRW (ablation) | RRNet + RRW · Nature | PSNR: 25.49 -> 25.96 | Helped |
| Add RDNet reflection-location guidance to RRNet + RRW (ablation) | RRNet + RRW · Real | PSNR: 23.38 -> 23.82 | Helped |
| Add RDNet reflection-location guidance to RRNet + RRW (ablation) | RRNet + RRW · SIR2 | PSNR: 24.82 -> 25.45 | Helped |
| Add RDNet reflection-location guidance (ablation) | Zhang et al. · Nature | PSNR: 22.31 -> 23.02 | Helped |
| Add RDNet reflection-location guidance (ablation) | Zhang et al. · Real | PSNR: 20.16 -> 20.76 | Helped |
| Add RDNet reflection-location guidance (ablation) | Zhang et al. · SIR2 | PSNR: 23.07 -> 23.46 | Helped |
| Add RDNet reflection-location guidance (ablation) | ERRNet · Nature | PSNR: 22.57 -> 23.64 | Helped |
| Add RDNet reflection-location guidance (ablation) | ERRNet · Real | PSNR: 20.67 -> 21.15 | Helped |
| Add RDNet reflection-location guidance (ablation) | ERRNet · SIR2 | PSNR: 22.97 -> 23.57 | Helped |
Surprising findings
- Reflection remnants persist in transmission layers obtained by raw-space subtraction in prior pipelines.
- Dataset-level PSNR of the strongest prior method on Real (YTMT, 22.94) and DSRNet on SIR2 are competitive, showing gains are not uniform across benchmarks.
Limitations
- Fails or leaves artifacts in saturated reflection regions where transmission content is nearly invisible.
- Requires aligned pairs and video capture with black cloth, which may be limited to scenes where the camera side can be occluded.
- Gains on SIR2 over DSRNet are modest relative to the Real and Nature benchmarks.
Experimental setup
Datasets: RRW Real Nature SIR2
Benchmarks: Nature Real SIR2
Metrics: PSNR SSIM
Baselines: BDN FRS Zhang et al. ERRNet RMNet Kim et al. IBCLN YTMT LANet PNACR DSRNet
Builds on: ERRNet IBCLN LANet Sobel operator VGG19
Closest papers by meaning
- Polarized Reflection Removal With Perfect Alignment in the Wild (CVPR 2020 · similarity 0.89)
- Single Image Reflection Removal Beyond Linearity (CVPR 2019 · similarity 0.89)
- Dereflection Any Image with Diffusion Priors and Diversified Data (AAAI 2026 · similarity 0.88)
- FIRM: Flexible Interactive Reflection ReMoval (AAAI 2025 · similarity 0.87)
- Removing Reflections from RAW Photos (CVPR 2025 · similarity 0.87)
- PolarFree: Polarization-based Reflection-Free Imaging (CVPR 2025 · 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.