Uniform Clustering Based Unsupervised Domain Adaptation for Semantic Segmentation

Supervised neural networks have achieved great success in semantic segmentation. However, these models cannot generalize well to a new dataset with different distributions since there exists an apparent domain gap between two domains. To reduce the domain discrepancy, unsupervised domain adaptation (UDA) techniques have been proposed for crossdomain semantic segmentation, which transfers knowledge from labeled source domain to unlabeled target domain. Recently, mainstream UDA approaches exploit adversarial training to learn domain-invariant and discriminative feature representations. Nonetheless, these methods do not consider the margin-aware class structure, leading to the mismatched class, especially for small and dense objects. To adapt pixel-level class structures, we propose the structurally uniform clustering loss. Our structurally uniform clustering loss supervises the feature generator and classifier simultaneously, aiming at narrowing the intra-class gap of different objects while maximizing the inter-class distance by driving different classes towards uniform distributions, which forms margin-aware class clusters. Besides, we also introduce confidence semantic features to prevent the training process from being influenced by low-confidence pixels. Experimental results on benchmark datasets demonstrate the effectiveness of our proposed method.