Closed HWalkingMan closed 2 years ago
How do you obtain the low-resolution images?
For train data, I downloaded zip file from (here from lsun) , and edited code from here to unzip mdb file and crop images. I used code from here (line:425-430) to crop images. Finally, I got 29343 HR images. Then, I used resize function supported by opencv-python to downsmaple image to 32px with bicubic.
For test data, I downloaded zip file from here, and selected the pictures whose names are in meta_info_Cat100_GT. Then, I used function from opencv-python to resize images and convert the file format. Finally, I got 100 HR image and 100 LR image in PNG format.
The released model is for images downsampled by the MATLAB imresize
function. Therefore, if you want to use the pre-trained model, please downsample the images with MATLAB. Otherwise, you can retrain the model using the degradation you want.
@HWalkingMan Hi, I tried all the steps without opencv-resizing as what you did, but the train data looks so weird. Many images contain unrelated contents. Have you ever encouter this problem??
@ckkelvinchan Do you use the crop codes likes here(line:425-430) when processing the LSUN-CAT?
We directly use imresize
to generate the training data, as we want to make sure that the main object is not cropped out.
But it is true that the aspect ratio could be modified. Someone would first center crop with size s x s, where s is the length of the shorter edge, and resize it to the desired size.
Thanks a lot! I'll try it.
Hi kelvin, I have tested the model (released on here) with Cat Test dataset, but I have obtained very poor test results.
At the same time, the metrics are also very unsatisfactory and far from the results given in the paper.
Could you give me some suggestions? Should I train model again or try something else?