ZstGAN: An adversarial approach for Unsupervised Zero-Shot Image-to-image Translation

Image-to-image translation models have shown remarkable ability on transferring images among different domains. Most of existing work follows the setting that source domain and target keep same at training inference phases, which cannot be generalized to scenarios for translating an image from unseen another domain. In this work, we propose Unsupervised Zero-Shot Translation (UZSIT) problem, aims learn a model can translate samples domains are not observed during training. Accordingly, framework called ZstGAN: By introducing adversarial scheme, ZstGAN learns each with domain-specific feature distribution is semantically consistent vision attribute modalities. Then domain-invariant features disentangled shared encoder generation. We carry out extensive experiments CUB FLO datasets, results demonstrate effectiveness proposed method UZSIT task. Moreover, shows significant accuracy improvements over state-of-the-art zero-shot learning methods FLO.