Semantics-Preserved Domain Adaptation with Target Diverse Perturbation and Test Ensembling for Image Segmentation

Despite the remarkable advance in unsupervised cross-domain image segmentation, existing works suffer from two main limitations. They either ignore the semantics preservation when transferring knowledge from source domain to target domain, or ignore to fully explore the rich information of the large amount of unlabeled data in target domain, leading to a bias segmentation model. To address these issues, we propose a novel semantics-preserved cross-domain image segmentation method with a new diverse image perturbation in the target domain, improving the capacity and robustness of the model. We also propose an effective test-time ensembling strategy for a more confident prediction result.