Semi-supervised Domain Adaptation for Semantic Segmentation via Active Learning with Feature- and Semantic-Level Alignments

Unsupervised domain adaptation (UDA) is a popular technique to reduce the manual annotation cost in semantic segmentation. However, due to the absence of strong supervision in the target domain, UDA is prone to biasing the decision boundary towards the source domain. To alleviate this issue, this paper proposes a more effective semi-supervised domain adaptation (SSDA) method for semantic segmentation via active learning with feature- and semantic-level alignments. Specifically, active learning is utilized to select those samples with high diversity and uncertainty from the target domain for labeling. These selected data could provide reliable clues for domain transfer since they reveal the intrinsic distribution of the target domain as well as including hard samples at boundaries. Moreover, to better adapt the segmentation model from the source data to the labeled target data selected above, we propose a scheme based on both feature- and semantic-level domain alignments. The feature-level domain alignment imposes the distribution consistency between the Transformer features of the two domains by adversarial learning, which is a global alignment. In contrast, the semantic-level domain alignment optimizes the affinity and divergence of the semantic representations across domains via contrastive learning, which is a local alignment. These two alignments jointly bridge the domain gap from both the global and the local views, respectively. In addition, the pseudo labels of the unlabeled data are generated to expand the labeled data and further strengthen the cross-domain segmentation in a self-training manner. Extensive experiments on segmentation benchmarks demonstrate the effectiveness of our proposed method.