SLS-CVPR2022

Junghun Oh, Heewon Kim, Seungjun Nah, Cheeun Hong, Jonghyun Choi, Kyoung Mu Lee

This repository is a Pytorch implementation of the paper "Attentive Fine-Grained Structured Sparsity for Image Restoration" from CVPR2022. [arXiv]

If you find this code useful for your research, please consider citing our paper:

@InProceedings{Oh_2022_CVPR,
  author = {Oh, Junghun and Kim, Heewon and Nah, Seungjun and Hong, Cheeun and Choi, Jonghyun and Lee, Kyoung Mu},
  title = {Attentive Fine-Grained Structured Sparsity for Image Restoration},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {June},
  year = {2022}
}

Proposed Method

Results

Quantitative results Qualitative results

Dataset and Pre-trained models

For super-resolution, we use DIV2K dataset to train and validate a model. You can download it here

After training, we evaluate trained models with benchmark datasets (Set14 - Zeyde et al. LNCS 2010, B100 - Martin et al. ICCV 2001, and Urban100 - Huang et al. CVPR 2015). You can download them here.

Unpack the downloaded tar files and change the args.dir_data in super-resolution/src/option.py to the directory where the DIV2K and benchmark datasets are located.

Since our method is applied to pre-trained models, you should download them through link, make a directory mkdir super-resolution/pretrained, and place the downloaded models in the directory.

Usage

Clone this repository.

git clone https://github.com/JungHunOh/SLS_CVPR2022.git
cd SLS_CVPR2022

cd super-resolution/src

For training,

bash ./scripts/train_sls_carnX4.sh $gpu $target_budget  # Training on DIV2K

For test,

bash ./scripts/test_sls_carnX4.sh $gpu $exp_name    # Test on Set14, B100, Urban100

To see the computational costs (w.r.t MACs and Num. Params.) of a trained model,

bash ./scripts/compute_costs.sh $gpu $model_dir

Acknowledgment

Our implementation is based on the following repositories:

JungHunOh / SLS_CVPR2022