It seems like the operation of 256x256 random cropping on Vimeo90K HR images is conducted before the model training process, instead of using an online fashion. The codes do not give descriptions on how this is done.
Could you explain the details of data pre-processing about (1) where to crop 256*256 (are image corners avoided?) ; (2) how many cropped videos are generated from the original videos?
I am now generating LR-HR pairs on-the-fly, so each vimeo clip is fetched only once during one epoch. The PSNR results on Vid4 Y channel (25.7 reproduced) are far worse than reported (27.69) after 70 epochs. I wonder whether this is due to inadequate training epochs since only 1 cropped clip for each HR clip is used.
It seems like the operation of 256x256 random cropping on Vimeo90K HR images is conducted before the model training process, instead of using an online fashion. The codes do not give descriptions on how this is done.
Could you explain the details of data pre-processing about (1) where to crop 256*256 (are image corners avoided?) ; (2) how many cropped videos are generated from the original videos?
I am now generating LR-HR pairs on-the-fly, so each vimeo clip is fetched only once during one epoch. The PSNR results on Vid4 Y channel (25.7 reproduced) are far worse than reported (27.69) after 70 epochs. I wonder whether this is due to inadequate training epochs since only 1 cropped clip for each HR clip is used.