Image-to-image translation for cross-domain disentanglement (NeurIPS ’18)

In this research, we propose using an image-to-image translation architecture to learn an image representation that is disentangled across domains. The representation consists of three parts: a shared part containing information that is common to both domains, and two exclusive parts, which only represent those factors of variation that are particular to each domain. We enforce disentanglement through a combination of multiple losses and a new network component called cross-domain autoencoder. We demonstrate the properties of our disentangled representation on various tasks such as multi-modal image translation and cross-domain retrieval on both synthetic and realistic datasets.(project page+code)

Comments are closed.