Maybe you misunderstand the track.
This track is for the real world image SR and LR-HR image pairs are captured with DSLRs. Thus, LRs are not down-sampled from HRs.
For the convenience of training, we cropped LRs into patches for the training set; for the validation and testing, we will evaluate methods on the whole images. Thus LRs might be larger than 1000*1000 in resolution.
There is no need to up-sample images because any synthetic image degradation is not consistent with the settings of this Real SR track.
Posted by: pengxu @ May 17, 2020, 7:36 a.m.I understand LR-HR image pairs are captured by DLSRs. What I don't understand is that the size of the LR image in training set is 4 times smaller than the HR image so it means we need to learn a 4x SR model. Then for validation, it makes sense to use the trained model for 4x SR. If the validation set doesn't need to be up-sampled for 4x, it must have been pre-upsampled to match the size of the HR image, otherwise the model cannot be applied to the validation. So my question is whether the validation images are pre-upsampled 4x by unknown operators?
Posted by: ZhiSong_Liu @ May 17, 2020, 9:18 a.m.Validation images are NOT pre-upsampled by unknown operators.
All the images are collected in the same manner. As claimed, for the convenience of training, we just cropped LR images into PATCHES with the size of 192*192 for the training set; for the validation and testing, we DO NOT crop LR images into cropped and will evaluate methods on the WHOLE IMAGES. Thus LRs in the validation set are larger in the resolution.
Posted by: pengxu @ May 17, 2020, 1:12 p.m.