부산대학교 도서관

Open-set recognition is a problem in which classes that do not exist in the training data can be presented at test time. Existing methods mostly take a multitask approach that integrates N-class classification and self-supervised pretext tasks, and they detect outliers by examining the distance to each class center in the feature space. Instead of relying on the learning through reconstruction, this paper explicitly uses distance learning to obtain the feature space for the open-set problem. In addition, although existing methods concatenate features from multiple tasks to measure the abnormality, we calculate it in each task-specific space independently and merge the results later. In experiments, the proposed method partially outperforms the state-of-the-art methods with significantly fewer parameters.