Skip to content

README.md

Latest commit

 

History

History
124 lines (91 loc) · 3.67 KB

README.md

File metadata and controls

124 lines (91 loc) · 3.67 KB
Raw

EDA — Natural Image Shadow Removal (SR)

Official implementation of shadow removal for the paper:

Elucidating the Design Space of Arbitrary-Noise-Based Diffusion Models (EDA)
CVPR 2026

SR is Task 3 in EDA: it removes shadows using sample-dependent basis noise h = X_S − x_0 (EDA Case 2, η = 10), built on the ShadowFormer backbone.

Acknowledgement

This module extends the ShadowFormer (AAAI 2023) codebase with EDA generalized diffusion training and multi-step deterministic sampling. We thank the ShadowFormer authors for releasing their implementation.

The overall diffusion scheduling follows the DDIM codebase used in other EDA tasks.

Method Summary (Paper Settings)

Item Setting
Basis Sample-dependent h = X_S − x_0
EDA case Case 2 — sample-dependent basis
η (training noise) 10: N = (10 + ε) / 11 · h, ε ~ N(0,1)
Forward process x_t = x_0 + √(1 − ᾱ_t) · N
Mask conditioning mask_t = m · √(1 − ᾱ_t) (Appendix E.4)
Training target ShadowFormer predicts x_0 directly
Loss L_SR = 𝔼 ‖x_θ(x_t, t) − x_0‖² (Charbonnier in code)
Diffusion steps T 100 (linear β: 0.0001 → 0.02)
Input size 320 × 320 × 3
Optimizer AdamW (ShadowFormer default), lr = 2×10⁻⁴
Training 2000 epochs, batch size 4
Sampling 5-step deterministic Euler from shadow image X_S

Repository Layout

shadow/
├── train_df.py          # EDA + ShadowFormer training
├── test_df.py           # 5-step deterministic inference
├── options.py           # Training hyperparameters
├── model.py             # ShadowFormer architecture
├── dataset.py           # ISTD / ISTD+ data loader
├── losses.py
├── utils/               # Data loading, checkpoints, metrics
├── warmup_scheduler/    # LR warmup (ShadowFormer official)
├── arg.txt
└── evaluation/          # MATLAB metrics (measure_shadow.m)

Requirements

This module extends ShadowFormer. Install Python, PyTorch, and other dependencies according to the baseline repository:

Baseline GitHub: https://github.com/GuoLanqing/ShadowFormer

A local requirements.txt is included for reference and matches the ShadowFormer stack.

Data Preparation

Use the ISTD dataset layout:

ISTD_Dataset/
├── train/
│   ├── train_A/   # shadow images (X_S)
│   ├── train_B/   # shadow masks
│   └── train_C/   # shadow-free GT (x_0)
└── test/
    ├── test_A/
    ├── test_B/
    └── test_C/

Usage

Train (EDA)

python train_df.py --train_dir ./ISTD_Dataset/train --val_dir ./ISTD_Dataset/test

Sample / Evaluate

python test_df.py \
  --input_dir ./ISTD_Dataset/test \
  --weights ./log/ShadowFormer_istd/models/model_best.pth \
  --times 5 \
  --save_images \
  --cal_metrics

See arg.txt for additional examples.

Citation

@inproceedings{qiu2026eda,
  title={Elucidating the Design Space of Arbitrary-Noise-Based Diffusion Models},
  author={Qiu, Xingyu and Yang, Mengying and Ma, Xinghua and others},
  booktitle={CVPR},
  year={2026}
}

@article{guo2023shadowformer,
  title={ShadowFormer: Global Context Helps Image Shadow Removal},
  author={Guo, Lan and Huang, Song and Liu, Dong and Cheng, Hao and Wen, Bihan},
  journal={AAAI},
  year={2023}
}