Research map / Detection and segmentation / CNN features
Recasting Regional Lighting for 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 a two-stage shadow removal method that decomposes shadow images into reflectance and illumination layers, then uses a diffusion-based local lighting correction and illumination-guided texture restoration, and annotates pixel-wise shadow masks for SRD.
Problem
Single-image shadow removal, where pixel-level color mappings jointly and implicitly model lighting and object textures, leaving shadow remnants in homogeneous regions and attenuated textures in textured regions.
Approach
Explicitly decompose shadow images via retinex-style self-supervised constraints into illumination and reflectance layers, regenerate homogeneous local illumination with a conditional local diffusion model, and restore textures in the reflectance layer conditioned on the corrected illumination through local and non-local lighting-to-texture attention.
Key results
- On SRD with manual masks, RMSE in shadow regions is 5.49 vs 7.11 for BMNet (best prior) and 7.45 for SGNet.
- On ISTD, shadow-region RMSE is 6.54 versus 7.60 for BMNet, a 13.9% reduction.
- On ISTD+, the method reports all-image PSNR 34.96 and shadow-region RMSE 5.69 versus 6.10 for BMNet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed shadow removal method instead of BMNet | shadow removal · SRD test, all regions | RMSE: 4.18 -> 3.66 | Helped |
| use the proposed shadow removal method instead of BMNet | shadow removal · ISTD test, all regions | RMSE: 5.02 -> 3.91 | Helped |
| use the proposed shadow removal method instead of BMNet | shadow removal · ISTD+ test, all regions | RMSE: 3.50 -> 2.87 | Helped |
| use local rather than global denoising (ablation) | LLC on ISTD · ISTD test, shadow regions | RMSE: 10.43 -> 9.40 | Helped |
| add the shadow mask as a conditioning input (ablation) | LLC on ISTD · ISTD test, shadow regions | RMSE: 9.40 -> 8.57 | Helped |
| add time-embedded non-shadow lighting condition C_t (ablation) | LLC on ISTD · ISTD test, shadow regions | RMSE: 8.57 -> 6.54 | Helped |
| add the local lighting-to-texture correspondence (ablation) | IGTR on ISTD · ISTD test, shadow regions | RMSE: 7.27 -> 6.54 | Helped |
| add the non-local lighting-to-texture correspondence (ablation) | IGTR on ISTD · ISTD test, shadow regions | RMSE: 7.00 -> 6.54 | Helped |
| replace standard self-attention with IGTR (ablation) | texture restoration on ISTD · ISTD test, shadow regions | RMSE: 7.32 -> 6.54 | Helped |
| replace input-level concatenation with IGTR (ablation) | texture restoration on ISTD · ISTD test, shadow regions | RMSE: 7.30 -> 6.54 | Helped |
| replace feature-level concatenation with IGTR (ablation) | texture restoration on ISTD · ISTD test, shadow regions | RMSE: 7.20 -> 6.54 | Helped |
| use bounding boxes rather than ground-truth masks (ablation) | proposed method on ISTD · ISTD test, shadow regions | RMSE: 5.98 -> 6.70 | Hurt |
Surprising findings
- Using only a bounding box instead of a pixel-wise shadow mask yields results that still outperform BMNet with masks on ISTD in shadow-region RMSE.
Limitations
- Diffusion-based local lighting correction adds parameters and inference time (1.7 s per 640x480 image).
- Evaluated on shadow removal benchmarks at 256 resolution only.
Experimental setup
Datasets: SRD ISTD ISTD+
Benchmarks: SRD ISTD ISTD+
Metrics: RMSE PSNR SSIM PER BER
Baselines: ST-CGAN DSC DHAN P+M+D DCGAN SP+M+I G2R AEF EMNet BMNet SGNet DeepUPE
Builds on: retinex theory DDPM BMNet co-attention SGNet
Closest papers by meaning
- OmniSR: Shadow Removal Under Direct and Indirect Lighting (AAAI 2025 · similarity 0.89)
- When Shadow Removal Meets Intrinsic Image Decomposition: A Joint Learning Framework Using Unpaired Data (AAAI 2025 · similarity 0.88)
- Polarization Guided Mask-Free Shadow Removal (AAAI 2025 · similarity 0.88)
- MetaShadow: Object-Centered Shadow Detection, Removal, and Synthesis (CVPR 2025 · similarity 0.86)
- Diff-Shadow: Global-guided Diffusion Model for Shadow Removal (AAAI 2025 · similarity 0.86)
- Detail-Preserving Latent Diffusion for Stable Shadow Removal (CVPR 2025 · similarity 0.86)
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.