Downsampling after alignment

[tristan-mcrae-rochester]

After you align the interpolated LR image with the HR ground truth image, what method of downsampling do you use to get the low resolution inputs down to 1/4 the resolution of the high resolution outputs as required by the SRGAN architecture?

[ngchc]

We adopt the bicubic method for interpolation as described in our paper. To adapt the dataset with SRGAN, we still adopt it for re-sampling. Generally, for SR methods pursuing the perceptual quality, it is not a mandatory requirement to feed inputs in low-resolution format.

[suke27]

as you said, you have HR result, and downsample using you methods. in you experiment, you may upsample the downsample images firstly, then do restruct using deeplearning model. am i right?

[ngchc]

@suke27 Hi. I am not quite sure if I have got your meaning. In the test phase, given an input image, a pre-interpolation (i.e., bicubic) is needed for VDSR while not for SRGAN. In the training phase, a re-sampling on the provided dataset is needed for SRGAN while not for VDSR.