Dear Organizer,
Does this contest mean we should process (crop then resize) the given high-resolution training images by our own at training? And at validation and testing, should we use trained network to process 1536 * 2048 images directly and upload the result? Or should we resize it to (W/4 * H/4) then recover it?
Thanks for your reply.
Regards,
zyz987
Hi, thank you for your question. The goal of this challenge is to learn to super-resolve images in the "source domain", i.e. the train-x and valid-x set images. These images have the same type of low-level artifacts. The output images should have a clean appearance, similar to those in the target domain set. You can train your network in any way you wish using the train-x set, e.g. process the images directly or just train on small crops. The performance is however measured on the images in the validation and testing set. These should be processed in full. You can check the resolution of these images by downloading them.
Regards,
Martin