Rethinking Exposure Correction for Spatially Non-uniform Degradation

This repository is the official implementation of "Rethinking Exposure Correction for Spatially Non-uniform Degradation"

Ao Li^1#, Jiawei Sun^1#, Le Dong^1*, Zhenyu Wang², Weisheng Dong¹

¹School of Artificial Intelligence, Xidian University

²Hangzhou Institute of technology, Xidian University

*: Corresponding Author.

#: Equal Contribution.

• Install the conda environment

conda create -n rethinkingEC python=3.10
conda activate rethinkingEC

*Install Pytorch

# CUDA 11.8
pip install torch==2.0.0+cu118 torchvision==0.15.1+cu118 --extra-index-url [https://download.pytorch.org/whl/cu118](https://download.pytorch.org/whl/cu118)

*Install other requirements

pip install -r requirements.txt

Please refer to the link below to download the dataset.Then remember to modify the dataset options in OPTIONS accordingly.

• LCDP
• MSEC
• SICE
• REED

• Testing. The testing configuration is in options/test/. Please put all the pre-training weights in the ./experiments. After running, the results of the visualization will be saved in . /results.

python test.py -opt options/test/test_lcdp.yml
python test.py -opt options/test/test_msec.yml
python test.py -opt options/test/test_sice.yml
python test.py -opt options/test/test_reed.yml

• Training. The training configuration is in options/train/.

python train.py -opt options/train/train_lcdp.yml
python train.py -opt options/train/train_sice.yml
python train.py -opt options/train/train_msec.yml

We provide the pretrained checkpoints in Baidu. Download and place them into ./experiments/ directory. The expected structure is:

RSNet/experiments/
├── MSEC_weights.pth       
├── LCDP_weights.pth
├── SICE_weights.pth  
└── REED_weights.pth       

Please cite the following paper if you feel our work useful to your research:

@misc{li2026rethinkingexposurecorrectionspatially,
      title={Rethinking Exposure Correction for Spatially Non-uniform Degradation}, 
      author={Ao Li and Jiawei Sun and Le Dong and Zhenyu Wang and Weisheng Dong},
      year={2026},
      eprint={2604.04136},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.04136}, 
}