COCM: Co-Occurrence-Based Consistency Matching in Domain-Adaptive Segmentation

This paper focuses on domain adaptation in a semantic segmentation task. Traditional methods regard the source and target as whole, image matching is determined by random seeds, leading to low degree of consistency between domains interfering with reduction gap. Therefore, we designed two-step, three-level cascaded strategy—co-occurrence-based (COCM)—in which two steps are: Step 1, design strategy from perspective category existence filter sub-image set highest whole domain, 2, which, spatial existence, propose method measuring PIOU score quantitatively evaluate co-occurring categories select best-matching image. The three levels mean that order improve importance low-frequency process, divide into according frequency co-occurrences domains; these are head, middle, tail levels, priority given categories. proposed COCM maximizes category-level has been proven be effective reducing gap while being lightweight. experimental results general datasets can compared those state-of-the-art (SOTA) methods.