Representation disentanglement aims at learning interpretable representation, which can be manipulated to analyze cross-domain visual data. By advancing adversarial learning and disentanglement techniques, our model learns domain-invariant representation with the associated domain-specific ones. Through our technique, we tackle cross-domain image generation and recognition simultaneously.

1.Our technique factorizes latent features into disjoint features describing domain-invariant and specific information.
2.Our network integrates adversarial learning, feature disentanglement, and generative modules in a unified architecture.
3.Our derived domain-invariant feature allows transfer learning, while the domain-specific feature enables diverse image generation across multiple domains.

The applications of our technique are related to transfer learning scenarios, and the most potential application is self-driving car. Through our technique, we are able to transfer the knowledge learning from one city or even simulated data to another city without labeled training data to reduce the expensive annotating cost and benefit to the industries devoted to self-driving car.