FogRemoval [ACCV'2022]

Introduction

Structure Representation Network and Uncertainty Feedback Learning for Dense Non-Uniform Fog Removal Asian Conference on Computer Vision (ACCV'2022)

[Paper] [Supp] [Poster] [Slides]

Datasets

1. SMOKE Dataset

The Fog Machine Antari W-515D

${FogRemoval}
|-- Dataset_day
    |-- [Smoke](https://www.dropbox.com/home/badweather/ACCV2022_defog/Dataset_day/Smoke)
      |-- train (110 pairs)
         |-- hazy  
         |-- clean
      |-- [test] (12 pairs) 
         |-- hazy  
         |-- clean  

[SMOKE Train] [SMOKE Test] [Ours Results]

[Dense-HAZE] [NH-HAZE] [O-HAZE]

Pre-trained Model

Download the pre-trained NH-HAZE model, put in results/NH-HAZE/model/NH-HAZE_params_0100000.pt

Test

python main_test.py --datasetpath [path_to_NH-HAZE dataset]

2. Fog Cityscapes

${FogRemoval}
|-- Dataset_day
    |-- Cityscapes
      |-- disparity 
      |-- leftImg8bit 
      |-- train (2,975 pairs)
         |-- hazy
         |-- clean 
      |-- test (1,525 pairs)
         |-- hazy  
         |-- clean 
      |-- generate_haze_cityscapes.m

Run the Matlab code to generate Synthetic Fog Cityscapes pairs:

Cityscapes/generate_haze_cityscapes.m

If smoke data is useful for your research, please cite our paper.

@InProceedings{Jin_2022_ACCV,
    author    = {Jin, Yeying and Yan, Wending and Yang, Wenhan and Tan, Robby T.},
    title     = {Structure Representation Network and Uncertainty Feedback Learning for Dense Non-Uniform Fog Removal},
    booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)},
    month     = {December},
    year      = {2022},
    pages     = {2041-2058}
}