peterzs / hardgan

HardGAN: A Haze-Aware Representation Distillation GAN for Single Image Dehazing

News

[2020/11/05] Our code released.

Reuqirements

Pytorch >= 1.1.0

numpy >= 1.6.0

tensorboardX

Quick start

• mkdir checkpoint data NH_results training_log
• Download datasets into ./data folder
• Use bash run.sh
• If you train on REDIES datasets, please use train_phrase 1.
• If you train on NTIRE2020， please train train_phrase 1 first. Then, copy trained trained parameter to G1 and G2(crop_size is [240, 240]). Finetune G2. Train G3 finally(crop_size is [960, 960]). You can choose whether to do data augmentation by aug_fog.py.

Citation

If your find our research is helpful for you, please cite our paper.

@inproceedings{deng2020hardgan,
  title={HardGAN: A Haze-Aware Representation Distillation GAN for Single Image Dehazing},
  author={Deng, Qili and Huang, Ziling and Tsai, Chung-Chi and Lin, Chia-Wen},
  booktitle={European Conference on Computer Vision},
  pages={722--738},
  year={2020},
  organization={Springer}
}