Discovering latent target subdomains for domain adaptive semantic segmentation via style clustering

Despite the great progress in unsupervised domain adaptation for semantic segmentation, most previous methods solely consider reducing the inter-domain gap caused by the distribution discrepancy between the source and target domain while not considering the sizeable intra-domain gap among the target domain itself due to the discrepancy among the target data. General intra-domain adaptation methods separate the target data into two splits based on how easily a sample can be segmented, which may not effectively capture the distributions within the target domain. In this paper, based on the observation that there exist diverse styles in the target samples, we propose a style clustering-based unsupervised domain adaptation method to separate the target data into subdomains iteratively. Since the target subdomain labels are unknown, we exploit multi-channel soft labels for adversarial training to close the intra-domain gap among these subdomains. In comparison with general intra-domain adaptation methods, our method can capture the latent distributions within the target data more sufficiently to close the intra-domain gap more effectively. The experiments of unsupervised domain adaptive segmentation tasks on benchmark datasets are conducted and the experimental results show the effectiveness of our method.

Automated summary

Previous methods in unsupervised domain adaptation for semantic segmentation fail to consider the intra-domain gap among the target domain itself. This paper proposes a style clustering-based unsupervised domain adaptation method, which effectively captures the latent distributions within the target data and reduces the intra-domain gap. Experimental results on benchmark datasets demonstrate the effectiveness of the proposed method.