eduzhai > Applied Sciences > Engineering >

Image-to-image Mapping with Many Domains by Sparse Attribute Transfer

  • Save

... pages left unread,continue reading

Document pages: 10 pages

Abstract: Unsupervised image-to-image translation consists of learning a pair ofmappings between two domains without known pairwise correspondences betweenpoints. The current convention is to approach this task with cycle-consistentGANs: using a discriminator to encourage the generator to change the image tomatch the target domain, while training the generator to be inverted withanother mapping. While ending up with paired inverse functions may be a goodend result, enforcing this restriction at all times during training can be ahindrance to effective modeling. We propose an alternate approach that directlyrestricts the generator to performing a simple sparse transformation in alatent layer, motivated by recent work from cognitive neuroscience suggestingan architectural prior on representations corresponding to consciousness. Ourbiologically motivated approach leads to representations more amenable totransformation by disentangling high-level abstract concepts in the latentspace. We demonstrate that image-to-image domain translation with manydifferent domains can be learned more effectively with our architecturallyconstrained, simple transformation than with previous unconstrainedarchitectures that rely on a cycle-consistency loss.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×