Awesome
Multi-Scale-Stage Network for Single Image Deblurring
Project | Paper | Supple
Official Implementation of ECCVW Paper
MSSNet: Multi-Scale-Stage Network for Single Image Deblurring
Kiyeon Kim, Seungyong Lee, Sunghyun Cho
POSTECH
ECCV 2022 Workshop (AIM 2022)
Architecture
Network Architecture
<img src="./img/mssnet.jpg" width="500">Training of MSSNet
<img src="./img/training.jpg" width="500">Installation
git clone https://github.com/kky7/MSSNet.git
To install warmup scheduler, refer MPRNet
cd pytorch-gradual-warmup-lr; python setup.py install; cd ..
Dependencies
- Python
- Pytorch 1.4 or 1.7
- scikit-image
- Tensorboard
Download
Dataset
Train
[GOPRO_Large]
Test
[Google Drive]
Pre-trained models
GOPRO
[Google Drive]
RealBlur
[Google Drive]
Training
if you use one gpu or multiple gpu using data parallel, run
sh sh_train_mssnet.sh
if you use multiple gpu using distributed data parallel, run
sh sh_train_mssnet_ddp.sh
Arguments
-
train_datalist : Text file with the path of image for training
-
val_datalist : Text file with the path of image for validation
-
checkdir : Path to save checkpoints
-
loadchdir : Path of checkpoint to load
-
data_root_dir : Root path of data in train_datalist
-
val_root_dir : Root path of data in val_datalist
-
isloadch : Whether to load checkpoint
-
isval : Whether to use validation
-
mgpu : Whether to use multiple gpu
-
wf, scale, vscale : Hyper-Parameters for channel size
Model wf scale vscale MSSNet_small 20 40 40 MSSNet 54 42 42 MSSNet_large 80 50 50
Testing
Run
sh sh_test_mssnet.sh
Arguments
- test_datalist : Text file with the path of image for testing
- data_root_dir : Root path of data in test_datalist
- load_dir : Path of checkpoint to load
- outdir : Path to save test results
- is_save : Whether to save test results
- is_eval : Whether to evaluate the model on the GoPro test dataset using psnr of skimage.metrics
Evaluation
Run evaluate_gopro.m
file to evaluate model on the gopro dataset.
This code is based on the MPRNet.
Acknowledgment
The code is based on the MPRNet, MIMO-UNet and ddp_example.
Citation
@inproceedings{Kim2022MSSNet,
author = {Kim, Kiyeon and Lee, Seungyong and Cho, Sunghyun},
title = {MSSNet: Multi-Scale-Stage Network for Single Image Deblurring},
booktitle = {Proc. of ECCVW (AIM)},
year = {2022}
}