This is the official PyTorch implementation of A Data-Centric Solution to NonHomogeneous Dehazing via Vision Transformer.
See more details in [report] , [paper], [certificate]

Our solution competes in NTIRE 2023 HR Non-homogeneous Dehazing Challenge, achieving the BEST performance in terms of PNSR, SSIM and LPIPS.

• (2023/6/18) We are the winner of the NTIRE 2023 HR Non-homogeneous Dehazing Challenge!
• (2022/6/01) We are invited to present our method at the NTIRE 2023 HR Non-homogeneous Dehazing Challenge.

• python3.7
• PyTorch >= 1.0
• NVIDIA GPU+CUDA
• numpy
• matplotlib
• tensorboardX(optional)

• Download ImageNet pretrained SwinTransformer V2 weights and our model weights.
• Download our dataset or the original dataset & gamma correction code

Please put train/val/test three folders into a root folder. This root folder would be the training dataset. Note, we have performed preprocessing to the data in folder train.

NTIRE2023_Val, and NTIRE2023_Test contain official validation and test. If you want to obtain val and test accuracy, please step towards the official competition server.

python train.py --data_dir data --imagenet_model SwinTransformerV2 --cfg configs/swinv2/swinv2_base_patch4_window8_256.yaml -train_batch_size 8 --model_save_dir train_result -train_epoch 6500

python test.py --imagenet_model SwinTransformerV2 --cfg configs/swinv2/swinv2_base_patch4_window8_256.yaml --model_save_dir ./output_img/test/best_result --ckpt_path ./checkpoints/best.pkl --hazy_data NTIRE2023_Test --cropping 4

• Using this command line for generating outputs of test data, the dehazed results could be found in: ./output_img/test/best_result
• This testing command line requires GPU memory >= 40 GB to ensure best results If GPU memory < 40 GB, please use " --cropping 6 " instead

Results on NTIRE 2023 NonHomogeneous Dehazing Challenge test data:

If you use the code in this repo for your work, please cite the following bib entries:

@inproceedings{liu2023data,
  title={A Data-Centric Solution to NonHomogeneous Dehazing via Vision Transformer},
  author={Liu, Yangyi and Liu, Huan and Li, Liangyan and Wu, Zijun and Chen, Jun},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={1406--1415},
  year={2023}
}