search

manuelfritsche / real-world-sr

[ICCVW 2019] PyTorch implementation of DSGAN and ESRGAN-FS from the paper "Frequency Separation for Real-World Super-Resolution". This code was the winning solution of the AIM challenge on Real-World Super-Resolution at ICCV 2019
MIT License
162 stars 37 forks source link

Question on TrainDataset in dsgan #2

Closed xiongdl closed 4 years ago

xiongdl commented 4 years ago

Hi Manuel Fritsche. Thanks for open source code. I have read your paper, and the figure 3 in original paper shows the structure to generate LR with natural characteristics. We can clearly see that the input of Gd is bicubic downsampled LR image, and the input of Dd is the output of Gd and source domain image. But when I see dsgan/train.py in detail, I find that only source domain images are used during the training of DSGAN. This is not consistent with figure 3 in original paper. Can you explain the reason?

manuelfritsche commented 4 years ago

Since the DSGAN model is trained in the "Same Domain" setting, the input images are simply the bicubically downscaled source images, while the "real" images are cropped versions of the same source images. You can read more about the setup used for training in section 4.1.