Awesome

Tackling the Ill-Posedness of Super-Resolution through Adaptive Target Generation

[Paper]

Dependency

• Python 3.6
• PyTorch 1.3

First things to do

1. Clone this repo.

git clone https://github.com/yhjo09/AdaTarget
cd AdaTarget

2. Download pretrained models and extract it.

unzip models.zip

Test - Isotropic Gaussian Kernels

1. Download a zip file containing the images of the Gaussian8 test set, and extract it.

unzip Gaussian8.zip

• The images were generated by following the instructions in the IKC repository.

2. Run.

python test_Iso.py

3. Output images will be in ./output/Iso.

Test - Random Kernels

1. Download DIV2KRK test set (from KernelGAN) and unzip the zip file.

unzip DIV2KRK.zip

2. Run.

python test_RK.py

3. Output images will be in ./output/RK.

Test - Bicubic Downsampling

1. Download a zip file contains the images of the Set5, Set14, BSDS100, Urban100, and Manga109 test sets, and extract it.

unzip Bicubic.zip

2. Run.

python test_Bic.py

3. Output images will be in ./output/Bic.

Train - Pretraining ATG

Soon

Train - Training SR Net with ATG

1. Prepare training images.

• Download DIV2K HR images (from https://data.vision.ee.ethz.ch/cvl/DIV2K) and unzip it to ./DIV2K/DIV2K_train_HR/*.png.
• Download DIV2K LR images and unzip it to ./DIV2K/DIV2K_train_LR_RK/*.png.
  • The LR images were generated by using https://github.com/assafshocher/BlindSR_dataset_generator with random kernels.

2. Prepare validation images.

• Download DIV2K HR images and unzip it to ./DIV2K/DIV2K_valid_HR/*.png.
• Download DIV2K LR images and unzip it to ./DIV2K/DIV2K_valid_LR_RK/*.png.

3. Specify the pretrained parameters of the localization network to variable PRETRAINED_LOCNET.

4. Start training.

python train_RK.py

• Tensorboard log will be stored in ./pt_log.
• Checkpoints and validation results will be saved in ./AdaTarget/.

BibTeX

@InProceedings{jo2021adatarget,
   author = {Jo, Younghyun and Oh, Seoung Wug and Vajda, Peter and Kim, Seon Joo},
   title = {Tackling the Ill-Posedness of Super-Resolution through Adaptive Target Generation},
   booktitle = {The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
   month = {June},
   year = {2021}
}

External codes from

• LPIPS
• RRDB (ESRGAN)