Validation image size in super-resolution models

I can’t figure out if I should use a cropped image when validation the model after each epoch or if I should validate the model using the image original size.

Taking DIV2K dataset, the images in general are very large in size but a good model like RDN trained their model on 32x32 image size. when they claimed achieving 38 PSNR, did they achieve it on the full original validation image size or in cropped validation image with 32x32.

I Tested the same implementation when I crop the image the model validation is not stable and it goes higher than 40 PSNR but when validating on the original size I achieve maximum 37.2 PSNR.

Could you help to understand here please ?

1 Like