Track1
train-x:
Flickr2K-1x-noise
but valid-x is:
DIV2K-4x-noise
One is on 1x, another one is on 4x.
Is it right? Are they in the same domain?
If I use bicubic downsampling for train-x, then train-x is: noise first and then downsample, but valid-x is: downsample first and then noise.
Hi. What you say is correct. Both the train-x and valid-x set can be considered the same domain. Because the DIV2K is very clean, the initial 4x downsampling does not significantly alter the low level characteristics of the image.
If you downsample the train-x, the resulting images will look very different from the original train-x and val-x images though, since this operation effectively "cleans" most of the corruptions. This is the major challenge of Real-World image super-resolution.
For more information on this, please refer to our AIM 2020 challenge paper https://arxiv.org/abs/1911.07783 or our dataset and method paper https://arxiv.org/abs/1909.09629.
Also, note that the added synthetic corruptions are not just standard "noise" (eg additive Gaussian iid).
Posted by: mdanelljan @ Dec. 26, 2019, 1:08 p.m.