PMAL: Open Set Recognition via Robust Prototype Mining

Open Set Recognition (OSR) has been an emerging topic. Besides recognizing predefined classes, the system needs to reject unknowns. Prototype learning is a potential manner handle problem, as its ability improve intra-class compactness of representations much needed in discrimination between known and In this work, we propose novel Mining And Learning (PMAL) framework. It prototype mining mechanism before phase optimizing embedding space, explicitly considering two crucial properties, namely high-quality diversity set. Concretely, set candidates are firstly extracted from training samples based on data uncertainty learning, avoiding interference unexpected noise. Considering multifarious appearance objects even single category, diversity-based strategy for filtering proposed. Accordingly, space can be better optimized discriminate therein classes Extensive experiments verify good characteristics (i.e., diversity) embraced mining, show remarkable performance proposed framework compared state-of-the-arts.