Collective Decision for Open Set Recognition

In open set recognition (OSR), almost all existing methods are designed specially for recognizing individual instances, even these instances collectively coming in batch. Recognizers decision either reject or categorize them to some known class using empirically-set threshold. Thus the threshold plays a key role. However, selection it usually depends on knowledge of classes, inevitably incurring risks due lacking available information from unknown classes. On other hand, more realistic OSR system should NOT just rest but go further, especially discovering hidden classes among whereas do not pay special attention. this paper, we introduce novel collective/batch strategy with an aim extend new discovery while considering correlations testing instances. Specifically, collective decision-based framework (CD-OSR) is proposed by slightly modifying Hierarchical Dirichlet process (HDP). Thanks HDP, our CD-OSR does need define and can implement simultaneously. Finally, extensive experiments benchmark datasets indicate validity CD-OSR.