Is it useful to use Siamese network structure for GANs like sharing latent space between generators in cGAN , or also with discriminators.

Thinking about it, like giving the generator tips about the knowledge-base of the discriminator, to target the problem of discriminator forgetting and increase the chance of convergence. Because than the discriminator prediction confidence is dependent of the generators construction (what’s anyway the case, but now on a system base ).

What do you think ? Didn’t see it in recent papers that often, just in this one, but it’s more like a pix2pix transformation, and just works so well, because they are using the segmentation masks of A, to get good segmentation results on B‘ ( A transformed to B). Didn’t find any approaches to sth like leaky discriminators.



You must log in to answer this question.

Browse other questions tagged .