AllWeatherNet

🏅 ICPR 2024 Best Paper Award

Unified Image Enhancement for Autonomous Driving under Adverse Weather and Lowlight Conditions

AllWeatherNet is an advanced image enhancement framework designed to improve the visibility of images captured in challenging weather conditions, such as snow, rain, fog, and low-light environments. This solution is particularly tailored for autonomous driving applications, ensuring safety and performance in diverse environments.

Key Features

• Unified Enhancement: Enhances images captured under various adverse weather conditions, including snowy, rainy, foggy, and nighttime scenarios.
• Scaled-Illumination Attention: Employs a robust scaled-illumination attention mechanism to maintain focus on the road across different conditions.
• Hierarchical Discrimination: Utilizes hierarchical patch-level discrimination at scene, object, and texture levels for more effective enhancement.

Environment Setup

To set up the required environment, please run:

conda create -n allweathernet python=3.8
conda activate allweathernet
pip install torch==1.8.0+cu111 torchvision==0.9.0+cu111 torchaudio==0.8.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txt

Demo Instructions

1. Download the Pretrained Model:
   Download the pretrained model from this link and place it in the checkpoints folder within the repository.

2. Set the Demo Image Path: You can put the images to be tested under the folder of test_data/testA. Or you can specify the path by setting the dataroot variable in script.py but your folder should contains testB with an image as a placeholder. Your test image can either be the original or a downsampled version from the original dataset.

3. Run the Script:
   Execute the script using the following command:

   python scripts/script.py --predict
   

Training (to be updated)

1. Download the dataset from the ACDC official website.
2. Organize the dataset in the following structure:

   ├── ACDC
   │   ├── trainA  # Contains adverse weather images
   │   └── trainB  # Contains normal weather images
   

Acknowledgements

This project is inspired by EnlightenGAN. We greatly appreciate the authors for their outstanding contributions.

Citation

If you find this work useful, please cite:

@inproceedings{qian2025allweather,
  title={AllWeather-Net: Unified Image Enhancement for Autonomous Driving Under Adverse Weather and Low-Light Conditions},
  author={Qian, Chenghao and Rezaei, Mahdi and Anwar, Saeed and Li, Wenjing and Hussain, Tanveer and Azarmi, Mohsen and Wang, Wei},
  booktitle={International Conference on Pattern Recognition},
  pages={151--166},
  year={2025},
  organization={Springer}
}

To-Do List

• Release test code
• Clean and refine training code
• Add more documentation and tutorials