DAST: Unsupervised Domain Adaptation in Semantic Segmentation Based on Discriminator Attention and Self-Training

Unsupervised domain adaption has recently been used to reduce the shift, which would ultimately improve performance of semantic segmentation on unlabeled real-world data. In this paper, we follow trend propose a novel method shift using strategies discriminator attention and self-training. The strategy contains two-stage adversarial learning process, explicitly distinguishes well-aligned (domain-invariant) poorly-aligned (domain-specific) features, then guides model focus latter. self-training adaptively improves decision boundary for target domain, implicitly facilitates extraction domain-invariant features. By combining two strategies, find more effective way shift. Extensive experiments demonstrate effectiveness proposed numerous benchmark datasets.