POC: Periodical Orthogonal Center Loss For Open Set Classification

When designing classification models, people usually do not assume that there will be unknown classes in the test set, which never appeared in the training set. However, this tricky situation is very common in practical applications. Such test conditions are called Open Set environments. Now, how to make models have the ability to identify unknown classes in the open environment has become a topic of great concern to researchers. In this paper, we follow up on previous research, which focusses on using orthogonal class centers to detect the unknown. We explain the reasons for the poor performance of the previous class center update strategy and propose using the orthogonal loss applied to the class centers to restrict the update direction. In addition, we use the multi-head attention layer for centers’ calculation to find suitable projection space adaptively. Experiments show that our method improves the performance of preceding orthogonal center methods.