Adverse Weather Image Translation with Asymmetric and Uncertainty-aware GAN (AU-GAN)

Official Tensorflow implementation of Adverse Weather Image Translation with Asymmetric and Uncertainty-aware GAN (AU-GAN)
Jeong-gi Kwak, Youngsaeng Jin, Yuanming Li, Dongsik Yoon, Donghyeon Kim and Hanseok Ko
British Machine Vision Conference (BMVC), 2021

Intro

Night → Day (BDD100K)

Rainy night → Day (Alderdey)

Architecture

Our generator has asymmetric structure for editing day→night and night→day. Please refer our paper for details

Envs

git clone https://github.com/jgkwak95/AU-GAN.git
cd AU-GAN

# Create virtual environment
conda create -y --name augan python=3.6.7
conda activate augan

conda install tensorflow-gpu==1.14.0   # Tensorflow 1.14
pip install --no-cache-dir -r requirements.txt

Preparing datasets

Night → Day
Berkeley DeepDrive dataset contains 100,000 high resolution images of the urban roads for autonomous driving.

Rainy night → Day
Alderley dataset consists of images of two domains, rainy night and daytime. It was collected while driving the same route in each weather environment.

Please download datasets and then construct them following ForkGAN

Training

# Alderley (256x512)
python main.py       --dataset_dir alderley
                     --phase train
                     --experiment_name alderley_exp
                     --batch_size 8 
                     --load_size 286 
                     --fine_size 256 
                     --use_uncertainty True

# BDD100k (512x1024)
python main.py       --dataset_dir bdd100k 
                     --phase train
                     --experiment_name bdd_exp
                     --batch_size 4 
                     --load_size 572 
                     --fine_size 512 
                     --use_uncertainty True

Test

# Alderley (256x512)
python main.py       --dataset_dir alderley
                     --phase test
                     --experiment_name alderley_exp
                     --batch_size 1 
                     --load_size 286 
                     --fine_size 256

# BDD100k (512x1024)
python main.py       --dataset_dir bdd100k
                     --phase test
                     --experiment_name bdd_exp
                     --batch_size 1 
                     --load_size 572 
                     --fine_size 512

Pretrained model : Night to Day

You can use a pretrained model for BDD100K dataset (size: 256x512)
First, download our pretrained model (zip file)
and then, unzip in folder --> ./check/your-exp-name/bdd100k_256/

Additional results

Please check more results in full paper (Arxiv)

Uncertainty map

Demo

Also check the demo code implemented by Katsuya Hyodo in Here.
It is optimized for ONNX, TFLite and other formats. Thanks to him and his community members for the amazing work!

Citation

If our code is helpful for your research, please cite our paper:

@InProceedings{kwak_adverse_2021},
  author = {Kwak, Jeong-gi and Jin, Youngsaeng and Li, Yuanming and Yoon, Dongsik and Kim, Donghyeon and Ko, Hanseok},
  title = {Adverse Weather Image Translation with Asymmetric and Uncertainty-aware GAN},
  booktitle = {British Machine Vision Conference (BMVC)},
  month = {November},
  year = {2021}
}

Acknowledgments

Our code is bulided upon the ForkGAN implementation.

zenonLEE/AU-GAN