Over-exposure Correction via Exposure and Scene Information Disentanglement

Pytorch implementation of the paper Over-exposure Correction via Exposure and Scene Information Disentanglement.

contact yuhuicao@pku.edu.cn

Usage

Data Preperation

In our experiments, outdoor images and portrait images are from Place365 dataset. To adjust the images exposure, we use the method proposed in LECRM. The numpy implementation of LECRM can be found in the file LECRM.py of this repository.

Train

To train the model, run the following command line in the source code directory. For calculating style loss, VGG19 model can be downloaded in here. You may set other parameters based on your experiment setting.

For disentanglement model, the exposure adjust process has be embedded into data_cfg.py and you can place your original data into your data directory to train the model:

python main.py -model dise -name experiment_name -phase train -data_root yourdataroot --dir_in yourdatadir    

For recovery model, you can run LECRM.py to generate overexposed images and run main.py to train the model:

python main.py -model reco -name experiment_name -phase train -data_root yourdataroot --dir_in overdir --dir_gt gtdir   

Citation

If you find the code helpful in your research or work, please kindly cite our paper.

@InProceedings{Cao_2020_ACCV,
    author    = {Cao, Yuhui and Ren, Yurui and Li, Thomas H. and Li, Ge},
    title     = {Over-exposure Correction via Exposure and Scene Information Disentanglement},
    booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)},
    month     = {November},
    year      = {2020}
}