Semantic-Rearrangement-based Hierarchical Alignment for domain generalized segmentation

Domain generalized semantic segmentation is an essential computer vision task, for which models only leverage source data to learn semantic segmentation towards generalizing to the unseen target domains. Previous works typically address this challenge by global style randomization or feature regularization. In this paper, we observe that different local semantic regions exhibit different visual characteristics from the source domain to the target domain. Thus, methods focusing on global operations find it hard to capture such regional discrepancies, failing to construct domain-invariant representations with consistency from the local to global level. Therefore, we propose Semantic-Rearrangement-based Hierarchical Alignment (SRHA) to overcome this problem. SRHA first incorporates a Semantic Rearrangement Module (SRM), with semantic region randomization to sufficiently enhance the diversity of the source domain. A Hierarchical Alignment Constraint (HAC) is subsequently proposed with the help of such diversity to establish the global-regional-local consistent domain-invariant representations. By aligning features across randomized samples with domain-neutral knowledge at multiple levels, SRHA provides a more robust way to handle the source-target domain gap. Extensive experiments indicate the superiority of SRHA over the current state-of-the-art works on various benchmarks.