Distance-Based Background Class Regularization for Open-Set Recognition

Wonwoo Cho, Jaegul Choo

Abstract: In open-set recognition (OSR), classifiers should be able to reject unknown-class samples while maintaining robust closed-set classification performance. To solve the OSR problem based on pre-trained Softmax classifiers, previous studies investigated offline analyses, e.g., distance-based sample rejection, which can limit the feature space of known-class data items. Since such classifiers are trained solely based on known-class samples, one can use background class regularization (BCR), which employs background-class data as surrogates of unknown-class data during training phase, to enhance OSR performance. However, previous regularization methods have limited OSR performance, since they categorized known-class data into a single group and then aimed to distinguish them from anomalies. In this paper, we propose a novel distance-based BCR method suitable for OSR, which limits the feature space of known-class data in a class-wise manner and then makes background-class samples located far away from the limited feature space. Instead of conventional Softmax classifiers, we use distance-based classifiers, which utilize the principle of linear discriminant analysis. Based on the distance measure used for classification, we design a novel regularization loss function that can contrast known-class and background-class samples while maintaining robust closed-set classification performance. Through our extensive experiments, we show that the proposed method provides robust OSR results with a simple inference process.