DIRL: Domain-Invariant Representation Learning for Generalizable Semantic Segmentation

Model generalization to the unseen scenes is crucial real-world applications, such as autonomous driving, which requires robust vision systems. To enhance model generalization, domain through learning domain-invariant representation has been widely studied. However, most existing works learn shared feature space within multi-source domains but ignore characteristic of itself (e.g., sensitivity domain-specific style). Therefore, we propose Domain-invariant Representation Learning (DIRL) for utilizes prior guide enhancement capability. The guidance reflects in two folds: 1) Feature re-calibration that introduces Prior Guided Attention Module (PGAM) emphasize insensitive features and suppress sensitive features. 2): whiting proposes Whiting (GFW) remove correlations are style. We construct suppresses effect style on quality correlation As a result, our method simple yet effective, can robustness various backbone networks with little computational cost. Extensive experiments over multiple generalizable segmentation tasks show superiority approach other methods.