Bidirectional Domain Mixup for Domain Adaptive Semantic Segmentation

Mixup provides interpolated training samples and allows the model to obtain smoother decision boundaries for better generalization. The idea can be naturally applied domain adaptation task, where we mix source target domain-mixed adaptation. However, extension of from classification segmentation (i.e., structured output) is nontrivial. This paper systematically studies impact mixup under adaptive semantic task presents a simple yet effective strategy called Bidirectional Domain (BDM). In specific, achieve in two-step: cut paste. Given warm-up trained any techniques, forward perform threshold-based out unconfident regions (cut). After then, fill-in dropped with other region patches (paste). doing so, jointly consider class distribution, spatial structure, pseudo label confidence. Based on our analysis, found that BDM leaves transferable by cutting, balances dataset-level distribution while preserving natural scene context pasting. We coupled proposal various state-of-the-art models observe significant improvement consistently. also provide extensive ablation experiments empirically verify main components framework. Visit project page code at https://sites.google.com/view/bidirectional-domain-mixup