Black-box model adaptation for semantic segmentation

Abstract

Model adaptation aims to transfer knowledge in pre-trained source models to a new unlabeled dataset. Despite impressive progress, prior methods always need to access the source model and develop data-reconstruction approaches to align the data distributions between target samples and the generated instances, which may raise privacy concerns from source individuals. To alleviate the above problem, we propose a new method in the setting of Black-box model adaptation for semantic segmentation, in which only the pseudo-labels from multiple source domain is required during the adaptation process. Specifically, the proposed method structurally distills the knowledge with multiple classifiers to obtain a customized target model, and then the predictions of target data are refined to fit the target domain with co-regularization. We conduct extensive experiments on several standard datasets, and our method can achieve promising results.