Original link to this project: PFFNet.

Paper can be download from PFFNet

PFFNet is a solution for the NTIRE2018 image defogging challenge (20.549db for Indoor and 20.230db for Outdoor), final results could be found at NTIRE2018. We won the six place and awarded "Honorable Mention Award".

Improved version was accepted by ACCV2018.

All pretrained models can be found at: Here

Using data_argument.py 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 image

Using default parameter to train

python train.py --cuda --gpus=4 --train=/path/to/train --test=/path/to/test --lr=0.0001 --step=1000

python test.py --cuda --checkpoints=/path/to/checkpoint --test=/path/to/testimages

If 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}
}