Abstract
Unsupervised domain adaptation for semantic segmentation aims to transfer the knowledge learned from a labeled synthetic source domain to an unlabeled real-world target domain. The main challenge lies in the difference between the two domains, i.e., the so-called “domain gap”. Although the two domains are supposed to share the same set of class labels, the semantics encoded by the source labels are not always consistent with those of the target data. Some recent efforts have been taken to explore the domain-specific semantics by conducting a within-domain adaptation using the predicted pseudo labels of the target data. The quality of the pseudo labels is therefore essential to the within-domain adaptation. In this paper, we propose a unified framework to progressively facilitate the adaptation towards the target domain. First, we propose to conduct the cross-domain adaptation through a novel source label relaxation. The relaxed labels offer a good trade-off between the source supervision and the target semantics. Next, we propose a dual-level self-regularization to regularize the pseudo-label learning and also to tackle the class-imbalanced issue in the within-domain adaptation stage. The experiment results on two benchmarks, i.e., GTA5 → Cityscapes and SYNTHIA → Cityscapes, show considerable improvement over the strong baseline and demonstrate the superiority of our framework over other methods.