Judging classless GAN performance

Hello I’ve been attempting to use conditional GAN to generate snow depth(snow topography) images by training cGANs(pix2pix) on snow depth images conditioned on no-snow topography images. I’ve tried finding ways of judging my results but most scores like the Inception score needs a class to create a score, which I don’t have when the image just contains snow depth values.

I was wondering if anyone knows about scores or ways to compare a real image with its generated fake. Like comparing a ground truth snow image with its generated fake by f.ex. pixel values.


1 Like

Maybe this could help: https://www.kaggle.com/c/gan-getting-started/overview/evaluation
This is a GAN competition and CycleGAN is used as an option. I think it is close to what you need

1 Like