Saliency Region Normalization Combined with Contrastive Learning for unsupervised Domain Adaptive Semantic Segmentation

By using labeled source domain data to train supervised models, unsupervised domain adapted (UDA) semantic segmentation tries to produce precise predictions in the unlabeled target domain. Previous methods failed to account for domain differences in class information, and poor alignment decreased generalizability in its place. Intending to optimize the quality of pseudo labels while learning features with distinct boundaries from the pixel, feature, and class perspectives, this paper proposes a Saliency Region Normalization (SRN) combined with Contrastive Learning (CL) for domain adaptation network. This method also establishes excellent cross-domain connections by learning invariant features, and, in the end, reduces domain gaps. According to the experimental findings, the GTA 5->Cityscapes and SYSTHIA->Cityscapes datasets yield 74.5% and 66.9% mIoU, respectively.