Maclory / Deep-Iterative-Collaboration

Pytorch implementation of Deep Face Super-Resolution with Iterative Collaboration between Attentive Recovery and Landmark Estimation (CVPR 2020)

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

The processing of image in HG model

LiangbinXie opened this issue · comments

Hi, when you train the HG model, what's the image processing? In my view, I think dividing 255 is necessaray, but do you substract the mean of images afterwards? If so, the mean of images is the same?

Hi, thank your for your interest in our work. You can find the the image processing steps in HRLandmarkDataset.py.
As for image normalization, you can find the mean here. We subtract the image with the mean RGB of whole CelebA dataset(0.509, 0.424, 0.378).

Thanks for your kind reply, but I still have some puzzle. I think you need to train the HG model in the first step, then train the DIC arch in the second step. So in the first step**(just FeedbackHourGlass arch)**, you mentioned that you follow this git repo, but what I found there about image processing is just dividing 255, with no image normalization. So I want to ask again whether you subtract the image with the mean RGB during the training of FeedbackHourGlass arch.
Sorry to bother you again, thanks!

I follow the network structure of the repo, but training setting is different, you can follow my own repo for face alignment with HourGlass network. From here you can see that I do subtract the mean.

OK, I see. Thanks for your kind reply~