Unknown Support Prototype Set for Open Set Recognition

Abstract

In real-world applications, visual recognition systems inevitably encounter unknown classes which are not present in the training set. Open set recognition aims to classify samples from known classes and detect unknowns, simultaneously. One promising solution is to inject unknowns into training sets, and significant progress has been made on how to build an unknowns generator. However, what unknowns exhibit strong generalization is rarely explored. This work presents a new concept called Unknown Support Prototypes, which serve as good representatives for potential unknown classes. Two novel metrics coined Support and Diversity are introduced to construct Unknown Support Prototype Set. In the algorithm, we further propose to construct Unknown Support Prototypes in the semantic subspace of the feature space, which can largely reduce the cardinality of Unknown Support Prototype Set and enhance the reliability of unknowns generation. Extensive experiments on several benchmark datasets demonstrate the proposed algorithm offers effective generalization for unknowns.