Latent Sparse Discriminative Learning for Face Image Set Classification

Image set classification has drawn much attention due to its promising performance to overcome various variations. Recently, the point-to-point distance-based methods have achieved the state-of-the-art performance by leveraging the distance between the gallery set and the probe set. However, there are two drawbacks that need to be defeated: 1) they do not fully exploit the discrimination that exists between different gallery sets; 2) they face the great challenge of high computational complexity as well as multi-parameters, usually caused by some obvious sparse constraints. To address these problems, this paper proposes a novel method, namely latent sparse discriminative learning (LSDL), for face image set classification. Specifically, a new term is proposed to exploit the relations between different gallery sets, which can boost the set discrimination so as to improve performance. Moreover, we use a latent sparse constraint to reduce the trade-off parameters and computational cost. Furthermore, an efficient solver is proposed to solve our LSDL. Experimental results on three benchmark datasets demonstrate the advantages of our propose.