Template-centric deep linear discriminant analysis for visual representation

In some real-world visual recognition tasks, instances are generated according to certain standards, which should serve as references during instance recognition. In this paper, we propose a template-centric representation learning (TCRL) framework that uses these standards as templates during recognition. The TCRL framework aims to learn a feature space where each instance is closely centered around its own template and away from the other templates. Within TCRL framework, we propose a template-centric objective function and a template-centric LDA layer, comprising two concrete models TDCNN and TDLDA. Experiments show that our method is superior to other traditional classification methods. The code will be made public after acceptance.