Awesome

AdaRevD: Adaptive Patch Exiting Reversible Decoder Pushes the Limit of Image Deblurring (CVPR2024)

Xintian Mao, Qingli Li and Yan Wang

Quick Run

Training

1. Download GoPro training and testing data
2. To train the main body of AdaRevD, download the pretrained model from NAFNet or UFPNet 百度网盘(提取码: xpjh), modify state_dict_pth_encoder in GoPro-AdaRevIDB-pretrain-4gpu.yml and run

cd AdaRevD
./train_4gpu.sh Motion_Deblurring/Options/GoPro-AdaRevIDB-pretrain-4gpu.yml

3. To train the classifier of AdaRevD, modify the pretrain_network_g in GoPro-AdaRevIDB-classify-4gpu.yml and run

./train_4gpu.sh Motion_Deblurring/Options/GoPro-AdaRevIDB-classify-4gpu.yml

Evaluation

To test the pre-trained models Google Drive or 百度网盘(提取码:dfce) on your own images (turn the 'pretrain' in yml from false to true for RevD), run

python Motion_Deblurring/val.py 

Results

Results on GoPro, HIDE, Realblur test sets: Google Drive or 百度网盘(提取码:27ex)

Citation

If you use , please consider citing:

@inproceedings{xintm2024AdaRevD, 
    title = {AdaRevD: Adaptive Patch Exiting Reversible Decoder Pushes the Limit of Image Deblurring},
    author = {Xintian Mao, Qingli Li and Yan Wang}, 
    booktitle = {Proc. CVPR}, 
    year = {2024}
    }

Contact

If you have any question, please contact mxt_invoker1997@163.com

Our Related Works

• Deep Residual Fourier Transformation for Single Image Deblurring, arXiv 2021. Paper | Code
• Intriguing Findings of Frequency Selection for Image Deblurring, AAAI 2023. Paper | Code
• LoFormer: Local Frequency Transformer for Image Deblurring, ACM MM 2024. Paper | Code

Reference Code:

• https://github.com/Fangzhenxuan/UFPDeblur
• https://github.com/megvii-research/NAFNet
• https://github.com/megvii-research/RevCol
• https://github.com/littlepure2333/APE
• https://github.com/INVOKERer/DeepRFT/tree/AAAI2023

Acknowledgment: This code is based on the BasicSR toolbox.