Awesome
MC-Blur: A Comprehensive Benchmark for Image Deblurring
Our propsoed MC-Blur Benchmark
We construct a large-scale multi-cause (MC-Blur) dataset for image deblurring. It consists of four blur types: uniform blurs, motion blurs by averaging continuous frames, heavy defocus blurs, and real-world blurs. We collect these images from more than 1000 diverse scenes such as buildings, city scenes, vehicles, natural landscapes, people, animals, and sculptures. MC-Blur Benchmark consits of four different subsets, i.e., Real high-fps based Motion-blurred subset (RHM), large-kernel UHD Motion-blurred subset (UHDM), large-scale heavy defocus blurred subset (LSD), and Real Mixed Blurry Qualitative subset (RMBQ).
Downloads
The images of the dataset can be downloaded from the links below.
Google Drive
<!-- - [RHM-250](https://drive.google.com/file/d/1hoCFNeP1GOszaJfLABBw35hCQQKWPMhV/view?usp=sharing) (6.2G for test) --> <!-- - [RHM-500](https://drive.google.com/file/d/13payJmIY6mssFXMSuOwf1aosA3n6bXb6/view?usp=sharing) (7.1G for test) --> <!-- - [RHM-1000](https://drive.google.com/file/d/1HJqV6Ogve-G6YGYJDH7MD5nKAQOTnVMG/view?usp=sharing) (9.3G for test) -->- RHM-250-500-1000 (117G total data)
- UHDM (278G total data)
- LSD (34G total data)
- RMBQ (110G total data)
Baidu Cloud (How to unzip?)
- RHM-250-500-1000 (117G total data) (password:ohzp)
- UHDM (278G total data) (password:p78n)
- LSD (34G total data) (password:sbtu)
- RMBQ (110G total data) (password:nwq8)
Download MC-Blur benchmark from the script, run
python download_data.py
Note: The above script will download all subsets of the MC-Blur. You can use "--data" to select. For example:
python download_data.py --data "UHDM_train_test"
Some visual examples of MC-Blur Dataset
Visual examples for each subset of our MC-Blur Dataset.
Some code steps in synthesizing dataset
See detail in README.
Benchmarking Study
Methods
Date | Publication | Title | Abbreviation | Code | Platform |
---|---|---|---|---|---|
2017 | CVPR | Deep multi-scale convolutional neural network for dynamic scene deblurring paper | DeepDeblur | Code | Pytorch |
2018 | CVPR | Deblurgan: Blind motion deblurring using conditional adversarial networks paper | DeblurGAN | Code | Pytorch |
2018 | CVPR | Scale-recurrent network for deep image deblurring paper | SRN | Code | Tensorflow |
2019 | ICCV | DeblurGAN-v2: Deblurring (Orders-of-Magnitude) Faster and Better paper | DeblurGAN-v2 | Code | Pytorch |
2019 | CVPR | Deep Stacked Hierarchical Multi-Patch Network for Image Deblurring paper | DMPHN | Code | Pytorch |
2020 | CVPR | Deblurring by Realistic Blurring paper | DBGAN | Code | Pytorch |
2021 | CVPR | Multi-Stage Progressive Image Restoration paper | MPRNet | Code | Pytorch |
2022 | CVPR | Restormer: Efficient Transformer for High-Resolution Image Restoration paper | Restormer | Code | Pytorch |
2021 | ICCV | Rethinking Coarse-To-Fine Approach in Single Image Deblurring paper | MIMO-UNet | Code | Pytorch |
Metrics
Abbreviation | Full-/Non-Reference | Platform | Code |
---|---|---|---|
PSNR (Peak Signal-to-Noise Ratio) | Full-Reference | ||
SSIM (Structural Similarity Index Measurement) | Full-Reference | MATLAB | Code |
NIQE (Naturalness Image Quality Evaluator) | Non-Reference | MATLAB | Code |
SSEQ (No-reference Image Quality Assessment Based on Spatial and Spectral Entropies) | Non-Reference | MATLAB | Code |
Results for 250-fps images from RHM Set
Method | PSNR | SSIM | Parameter |
---|---|---|---|
DeepDeblur | 30.38 | 0.8766 | 11.72 M |
DeblurGAN | 24.89 | 0.6364 | 6.07 M |
SRN | 30.57 | 0.8799 | 6.88 M |
DeblurGAN-v2 | 26.99 | 0.8061 | 7.84 M |
DMPHN | 30.42 | 0.8768 | 21.69 M |
DBGAN | 27.89 | 0.8191 | 11.59 M |
MPRNet | 31.52 | 0.9239 | 20.13 M |
Restormer | 30.41 | 0.9106 | 26.10 M |
MIMO-UNet | 32.02 | 0.9285 | 6.81 M |
Results for 500-fps images from RHM Set
Method | PSNR | SSIM | Parameter |
---|---|---|---|
DeepDeblur | 31.08 | 0.8974 | 11.72 M |
DeblurGAN | 24.66 | 0.6748 | 6.07 M |
SRN | 31.54 | 0.9051 | 6.88 MB |
DeblurGAN-v2 | 27.67 | 0.8320 | 7.84 M |
DMPHN | 31.43 | 0.9018 | 21.69 M |
DBGAN | 28.36 | 0.8388 | 11.59 M |
MPRNet | 32.08 | 0.9300 | 20.13 M |
Restormer | 30.98 | 0.9160 | 26.10 M |
MIMO-UNet | 32.89 | 0.9398 | 6.81 M |
Results for 1000-fps images from RHM Set
Method | PSNR | SSIM | Parameter |
---|---|---|---|
DeepDeblur | 32.41 | 0.8966 | 11.72 M |
DeblurGAN | 25.20 | 0.6535 | 6.07 M |
SRN | 32.69 | 0.0.9016 | 6.88 M |
DeblurGAN-v2 | 29.81 | 0.8461 | 7.84 M |
DMPHN | 32.41 | 0.9096 | 21.69 M |
DBGAN | 29.66 | 0.8318 | 11.59 M |
MPRNet | 33.36 | 0.9332 | 20.13 M |
Restormer | 32.77 | 0.9264 | 26.10 M |
MIMO-UNet | 33.75 | 0.9360 | 6.81 M |
Results on UHDM Set
Method | PSNR | SSIM | Parameter |
---|---|---|---|
DeepDeblur | 22.23 | 0.6322 | 11.72 M |
DeblurGAN | 20.39 | 0.5568 | 6.07 M |
SRN | 22.28 | 0.6346 | 6.88 M |
DeblurGAN-v2 | 21.03 | 0.5839 | 7.84 M |
DMPHN | 22.20 | 0.6378 | 21.69 M |
DBGAN | 21.52 | 0.6025 | 11.59 M |
MPRNet | 23.70 | 0.7472 | 20.13 M |
Restormer | 22.39 | 0.7356 | 26.10 M |
MIMO-UNet | 22.97 | 0.7317 | 6.81 M |
Results on LSD Set
Method | PSNR | SSIM | Parameter |
---|---|---|---|
DeepDeblur | 20.73 | 0.7218 | 11.72 M |
DeblurGAN | 20.04 | 0.6335 | 6.07 M |
SRN | 21.66 | 0.7664 | 6.88 M |
DeblurGAN-v2 | 21.13 | 0.6964 | 7.84 M |
DMPHN | 21.23 | 0.7519 | 21.69 M |
DBGAN | 21.56 | 0.7536 | 11.59 M |
MPRNet | 21.32 | 0.7897 | 20.13 M |
Restormer | 22.35 | 0.8072 | 26.10 M |
MIMO-UNet | 22.56 | 0.7985 | 6.81 M |
Citation
If you think this work is useful for your research, please cite the following paper.
@inproceedings{zhang2023benchmarking,
title={MC-Blur: A Comprehensive Benchmark for Image Deblurring},
author={Zhang, Kaihao and Wang, Tao and Luo, Wenhan and Chen, Boheng and Ren, Wenqi and Stenger, Bjorn and Liu, Wei and Li, Hongdong and Yang Ming-Hsuan},
booktitle={IEEE Transactions on Circuits and Systems for Video Technology},
year={2023}
}
License
The MC-Blur dataset is released under CC BY-NC-ND license.