Isobe Takashi ; Jia Xu ; Chen Shuaijun ; He Jianzhong ; Shi Yongjie ; Liu Jianzhuang ; Lu Huchuan ; Wang Shengjin

Abstract

Recently unsupervised domain adaptation for the semantic segmentation taskhas become more and more popular due to high-cost of pixel-level annotation onreal-world images. However, most domain adaptation methods are only restrictedto single-source-single-target pair, and can not be directly extended tomultiple target domains. In this work, we propose a collaborative learningframework to achieve unsupervised multi-target domain adaptation. Anunsupervised domain adaptation expert model is first trained for eachsource-target pair and is further encouraged to collaborate with each otherthrough a bridge built between different target domains. These expert modelsare further improved by adding the regularization of making the consistentpixel-wise prediction for each sample with the same structured context. Toobtain a single model that works across multiple target domains, we propose tosimultaneously learn a student model which is trained to not only imitate theoutput of each expert on the corresponding target domain, but also to pulldifferent expert close to each other with regularization on their weights.Extensive experiments demonstrate that the proposed method can effectivelyexploit rich structured information contained in both labeled source domain andmultiple unlabeled target domains. Not only does it perform well acrossmultiple target domains but also performs favorably against state-of-the-artunsupervised domain adaptation methods specially trained on a singlesource-target pair