Our solution for NTIRE2018 Image Dehazing Challenge (20.549db for Indoor and 20.230db for Outdoor), final results could be refer at NTIRE2018. Futher version is accepted by ACCV2018 https://arxiv.org/pdf/1810.02283.pdf. All pretrained models can be found at Here
Using data_argument to enchance the datasets, it will produce below datasets
$ python dara_argument.py --fold_A=IndoorTrainHzay --fold_B=IndoorTrainGT --fold_AB=IndoorTrain
IndoorTrain
\data hazy image
\label clear imageUsing default parameter to train
python train.py --cuda --gpus=4 --train=/path/to/train --test=/path/to/test --lr=0.0001 --step=1000python test.py --cuda --checkpoints=/path/to/checkpoint --test=/path/to/testimagesIf you use the code in this repository, please cite our paper:
@inproceedings{mei2018pffn,
title={Progressive Feature Fusion Network for Realistic Image Dehazing},
author={Mei, Kangfu and Jiang, Aiwen and Li, Juncheng and Wang, Mingwen},
booktitle={Asian Conference on Computer Vision (ACCV)},
year={2018}
}