Kernel-Aware Burst Blind Super-Resolution, WACV 2023

[paper]

Dependencies

• OS: Ubuntu 18.04
• Python: Python 3.7
• nvidia :
  • cuda: 10.1
  • cudnn: 7.6.1
• Other reference requirements

Quick Start

1.Create a conda virtual environment and activate it

conda create -n pytorch_1.6 python=3.7
source activate pytorch_1.6

2.Install PyTorch and torchvision following the official instructions

conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.1 -c pytorch

3.Install build requirements

pip3 install -r requirements.txt

4.Install DCN

cd DCNv2-pytorch_1.6
python3 setup.py build develop # build
python3 test.py # run examples and check

Training

# Modify the root path of training dataset and model etc.
python main.py --n_GPUs 4 --lr 0.0002 --decay 100-200 --save kbnet --model KBNet --n_feats 128 --n_resblocks 8 --n_resgroups 5 --batch_size 32 --burst_size 8 --patch_size 256 --scale 4 --loss 1*L1

Testing

1. Data preparation

# Modify the output path of test dataset in make_validation_set.py - root = '../test_set_3.2_4.8'
python make_validation_set.py

2. Run models on test dataset:

# Modify the path of test dataset and the path of the trained model
python test.py --root ../test_set_0_1.6 --n_GPUs 1 --model KBNet --n_feats 128 --n_resblocks 8 --n_resgroups 5 --batch_size 64 --burst_size 2 --scale 4 --pre_train ../train_log/KBNet/real_models/kbnet/KBNetbest_epoch.pth

Results

Citations

If EBSR helps your research or work, please consider citing EBSR. The following is a BibTeX reference.

@article{lian2021kernel,
  title={Kernel-aware Burst Blind Super-Resolution},
  author={Lian, Wenyi and Peng, Shanglian},
  journal={arXiv preprint arXiv:2112.07315},
  year={2021}
}

Acknowledgments

This repo is based on EBSR and IKC.