Learning Deep Features via Congenerous Cosine Loss for Person Recognition

Person recognition aims at recognizing the same identity across time and space with complicated scenes and similar appearance. In this paper, we propose a novel method to address this task by training a network to obtain robust and representative features. A key observation is that traditional cross entropy loss only enforces the inter-class variation among samples and ignores to narrow down the similarity within each category. We propose a congenerous cosine loss to enlarge the inter-class distinction as well as alleviate the inner-class variance. Such a design is achieved by minimizing the cosine distance between sample and its cluster centroid in a cooperative way. Our method differs from previous work in person recognition that we do not conduct a second training on the test subset and thus maintain a good generalization ability. The identity of a person is determined by measuring the similarity from several body regions in the reference set. Experimental results show that the proposed approach achieves better classification accuracy against previous state-of-the-arts.