PDA: Progressive Domain Adaptation for Semantic Segmentation

Abstract

The unsupervised domain adaptation semantic segmentation task is challenging due to the distribution shift problem between the source and the target domains. In this paper, we provide a novel perspective to address this problem. Currently, in the literature, it is shown that output-level domain adaptation networks (OL-DAN) can generate outputs with smaller distribution shift. Motivated by this phenomenon, a Progressive Domain Adaptation (PDA) framework is proposed, which uses the outputs generated by OL-DAN as the auxiliary domain images to progressively align the distribution between the source and target domains. Unlike existing two-stage input-level domain adaptation methods which use an image translation network as a standalone model to generate the auxiliary domain images, the PDA is an end-to-end framework that contains an OL-DAN and a domain fusion domain adaptation network (DF-DAN). The OL-DAN aims to gradually generate the outputs with smaller distribution shift and more accurate semantic structures as the auxiliary domain images in every iteration optimization. The DF-DAN is proposed to further mine the domain-invariant information from the auxiliary images and then fuse the features learned from the original images and the auxiliary images to obtain a richer representation. Finally, the distribution between the source and target domains is aligned to optimize the OL-DAN and DF-DAN. Experiments demonstrate that the proposed PDA achieves superior performance on three cross-domain semantic segmentation benchmarks.