Intra and Inter Class Consistency Domain Adaptation for Semantic Segmentation

We consider the problem of unsupervised domain adaptation for the task of semantic segmentation. Existing methods mainly focus on matching the marginal distributions between two domains from the global level through adversarial learning. However, this global alignment method does not consider the class-level joint distribution, which will increase the domain invariance of class features but weaken the discriminability. In this paper, we propose a novel category-level domain adaptation network for semantic segmentation, which explicitly reduces the divergence of intra-class features and enhances the separability of inter-class features. Extensive experiments verify the effectiveness of our proposed IaICC model on challenging unsupervised domain adaptation tasks, i.e., GTA5-Cityscapes.