Research map / Architectures and efficiency / Attention and transformers
HomoFormer: Homogenized Transformer for Image 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 HomoFormer, a local-window transformer that randomly shuffles pixels (with exact inverse) so shadow degradation becomes spatially homogeneous before local self-attention, plus a depth-wise-conv FFN (SMLP), for image shadow removal.
Problem
Shadow degradation is spatially non-uniform and diverse, which conflicts with the weight-sharing nature of CNNs and local window transformers, leading to compromised shadow removal; global attention is too costly at high resolution.
Approach
Randomly shuffle pixels across the whole image (and invert the shuffle afterwards) to homogenize the spatial distribution of shadow degradation so that weight-sharing local self-attention can handle it; structure is modeled in the FFN via depth-wise convolutions instead of positional encoding in attention.
Key results
- On SRD, HomoFormer reaches 35.37 dB PSNR and 3.33 MAE versus ShadowFormer's 32.90 dB and 4.04 MAE.
- On ISTD+, HomoFormer achieves overall MAE 2.7 (HomoFormer+ 2.6), beating ShadowFormer and ShadowDiffusion (both 2.8 and 2.7 respectively, with ShadowDiffusion tied at 2.7).
- Removing random shuffle raises ISTD+ shadow-region MAE from 5.0 to 6.4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add random shuffle and inverse shuffle to local self-attention (ablation) | HomoFormer · ISTD+ | MAE: 6.4 -> 5.0 (shadow); 2.5 -> 2.3 (non-shadow); 3.0 -> 2.7 (all) | Helped |
| Add depth-wise convolution for structure modeling in the FFN (ablation) | HomoFormer · ISTD+ | MAE: 5.6 -> 5.0 (shadow); 2.5 -> 2.3 (non-shadow); 2.9 -> 2.7 (all) | Helped |
| Use HomoFormer instead of ShadowFormer | SRD · SRD, all region | PSNR: 32.90 dB -> 35.37 dB | Helped |
| Use HomoFormer+ with Monte Carlo averaging instead of HomoFormer | HomoFormer · ISTD+ | MAE: 5.0 -> 5.0 (shadow); 2.3 -> 2.2 (non-shadow); 2.7 -> 2.6 (all) | Mixed |
Surprising findings
- Removing the GELU activation in the structure-modeling FFN is claimed to be redundant with element-wise multiplication and is validated by ablation.
- Variance across random shuffles is reported as negligible, and shuffle-induced variance can be used as an uncertainty estimate that predicts errors.
Limitations
- Monte Carlo averaging with 8 samples is needed for the best results, increasing inference cost.
- Evaluated only on two shadow removal benchmarks; generalization to other restoration tasks is untested.
Experimental setup
Datasets: ISTD+ (Adjusted ISTD) SRD
Benchmarks: ISTD+ SRD
Metrics: MAE (LAB) PSNR SSIM
Baselines: ShadowFormer ShadowDiffusion BMNet SG-ShadowNet Li et al. (ICCV'23) Fu et al. (CVPR'21) Jin et al. (ICCV'21) DHAN G2R DeshadowNet
Builds on: Swin Transformer Restormer-style UNet Charbonnier loss Kang et al. pixel shuffle regularization
Closest papers by meaning
- Diff-Shadow: Global-guided Diffusion Model for Shadow Removal (AAAI 2025 · similarity 0.87)
- Polarization Guided Mask-Free Shadow Removal (AAAI 2025 · similarity 0.87)
- Detail-Preserving Latent Diffusion for Stable Shadow Removal (CVPR 2025 · similarity 0.86)
- DeS3: Adaptive Attention-Driven Self and Soft Shadow Removal Using ViT Similarity (AAAI 2024 · similarity 0.84)
- HomoGen: Enhanced Video Inpainting via Homography Propagation and Diffusion (CVPR 2025 · similarity 0.84)
- When Shadow Removal Meets Intrinsic Image Decomposition: A Joint Learning Framework Using Unpaired Data (AAAI 2025 · similarity 0.83)
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.