search

yxgeee / FD-GAN

[NeurIPS-2018] FD-GAN: Pose-guided Feature Distilling GAN for Robust Person Re-identification.
https://yxgeee.github.io/projects/fdgan.html
281 stars 80 forks source link

question about image normalization #5

Closed xuxu116 closed 5 years ago

xuxu116 commented 5 years ago

The real images are normalized by mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], while the generated images are in range [-1, 1] because of the Tanh() function in generator. Is it necessary to consider this inconsistence when passing fake images to the Identity Discriminator? Or I just missing some details in your code. Thanks!

yxgeee commented 5 years ago

Yes, it is a minor issue in the origin version code. I have tried to normalize the generated images to have the same mean and std as real images, but found that it can slightly improve the generation results but have no influences on reID performance. Generation is a auxiliary task in this work, so I released the origin version code.

If you want to have better generation results, you can modify the tanh() to sigmoid(), and then apply the following normalization code to the generated images,

def norm_img(self, imgs, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]):
        mean = Variable(torch.Tensor(mean).view(1,3,1,1).expand_as(imgs)).cuda()
        std = Variable(torch.Tensor(std).view(1,3,1,1).expand_as(imgs)).cuda()
        imgs.sub_(mean).div_(std)
xuxu116 commented 5 years ago

Thanks for your reply!