Open-Set Hypothesis Transfer With Semantic Consistency

Unsupervised open-set domain adaptation (UODA) is a realistic problem where unlabeled target data contain unknown classes. Prior methods rely on the coexistence of both source and to perform alignment, which greatly limits their applications when are restricted due privacy concerns. In this paper we address challenging hypothesis transfer setting for UODA, from no longer available during domain. Specifically, propose use pseudo-labels novel consistency regularization data, using conventional formulations fails in setting. Firstly, our method discovers confident predictions performs classification with pseudo-labels. Then enforce model output consistent definite semantically similar transformed inputs, discovering all latent class semantics. As result, can be classified into discriminative classes coincided either or We theoretically prove that under perfect semantic transformation, proposed objective enforces recover information true labels prediction. Experimental results show outperforms state-of-the-art UODA benchmarks.