search

sanghyun-son / EDSR-PyTorch

PyTorch version of the paper 'Enhanced Deep Residual Networks for Single Image Super-Resolution' (CVPRW 2017)
MIT License
2.42k stars 668 forks source link

Training Data #99

Open miken1990 opened 5 years ago

miken1990 commented 5 years ago

Hi,

Thanks for the implementation and detailed explanation! I tried to understand from the paper/repo, but couldn't find an answer, what is the dataset that was used for training EDSR and MDSR? is it public?

Thanks, Michael

sanghyun-son commented 5 years ago

Hello.

We used a public dataset.

Please find word 'DIV2K' from README and you will get the answers you want.

Thank you!

miken1990 commented 5 years ago

Thanks for the answer!

1) As we understand from the article, you took each image in the DIV2K dataset, divided it to 48x48 patches (~ 1200 patches from each image) then flipped each patch with random horizontal and 90 degrees flips, so you ended up with around 7 Million images for training. Is that correct?

2) Our aim is to use the EDSR network to perform super resolution + artifact reduction and we are currently thinking about using a training set with a substantial size, but we are not sure that we have enough images, so we'll need to do augmentation. What is your opinion?

Thanks, Michael

m732367606 commented 5 years ago

Hi, do you know which part of the codes controls training dataset cropping stride? I wanna create nonoverlapping training datasets. Thank you!