Home

Awesome

Real-World Blur Dataset for Learning and Benchmarking Deblurring Algorithms

by Jaesung Rim, Haeyun Lee, Jucheol Won, Sunghyun Cho. [pdf] [project]

News

2022.10 Related work, Realistic Blur Synthesis for Learning Image Deblurring (real-world blur dataset and blur synthesis pipeline) is accepted by ECCV2022. Code and dataset are available at RSBlur github.

Result of RealBlur Test set

<img src="./imgs/qualatitive_result_web.png" width="100%" alt="Real Photo">

Installation

git clone --recurse-submodules https://github.com/rimchang/RealBlur.git

Prerequisites

We recommend virtual environment using conda or pyenv.

SRN-Deblur
DeblurGAN-v2

Download

For testing, download RealBlur.

For training same as our paper, download RealBlur, BSD-B, GoPro.

All datasets should be located in SRN-Deblur/testing_set/, SRN-Deblur/training_set/, DeblurGAN-v2/dataset/.

Also, we provide trained models. Please move checkpoint files to SRN-Deblur/checkpoints, DeblurGAN-v2/checkpoints.

Please check "link_file.sh" for appropriate linking of directories and files.

If you have a network problem, please use google drive link.

Training

# ./SRN-Deblur
python run_model.py --phase=train --batch=16 --lr=1e-4 --model=color --checkpoint_path=RealBlurJ_pre_trained+GOPRO+BSD500 --datalist=datalist/RealBlur_J_train_list.txt,datalist/BSB_B_Centroid_train.txt,datalist/datalist_gopro.txt --pre_trained=./checkpoints/color --load_iteration=523000 --warmup=1 --over_sampling=20000
python run_model.py --phase=train --batch=16 --lr=1e-4 --model=color --checkpoint_path=RealBlurR_pre_trained+GOPRO+BSD500 --datalist=datalist/RealBlur_R_train_list.txt,datalist/BSB_B_Centroid_train.txt,datalist/datalist_gopro.txt --pre_trained=./checkpoints/color --load_iteration=523000 --warmup=1 --over_sampling=20000

# ./DeblurGANv2
python train_RealBlur_J_bsd_gopro_pretrain_ragan_ls.py
python train_RealBlur_R_bsd_gopro_pretrain_ragan_ls.py

Testing

# ./SRN-Deblur
python run_model.py --phase=test --model=color --checkpoint_path=RealBlurJ_pre_trained+GOPRO+BSD500 --datalist=datalist/RealBlur_J_test_list.txt --height=784 --width=688
python run_model.py --phase=test --model=color --checkpoint_path=RealBlurR_pre_trained+GOPRO+BSD500 --datalist=datalist/RealBlur_R_test_list.txt --height=784 --width=688

# ./DeblurGANv2
python predict.py --img_pattern=./datalist/RealBlur_J_test_list.txt --weights_path=checkpoints/last_deblur_gan_v2_RealBlur_J_bsd_gopro_pretrain_ragan_ls_10000.h5
python predict.py --img_pattern=./datalist/RealBlur_R_test_list.txt --weights_path=checkpoints/last_deblur_gan_v2_RealBlur_R_bsd_gopro_pretrain_ragan_ls_10000.h5

Evaluation

# python3, skimage == 0.17.2, cv2==4.2.0.32
python evaluation_RealBlur_ecc.py --gt_root=dataset/RealBlur-J_ECC_IMCORR_centroid_itensity_ref --input_dir=RealBlur_J --core=1 

Post-processing

Please go to post processing

License

The RealBlur dataset is released under CC BY 4.0 license.

Citation

If you use our dataset for your research, please cite our paper.

@inproceedings{rim_2020_ECCV,
 title={Real-World Blur Dataset for Learning and Benchmarking Deblurring Algorithms},
 author={Rim, Jaesung and Lee, Haeyun and Won, Jucheol and Cho, Sunghyun},
 booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
 year={2020}
}