Kinship verification in multi-linear coherent spaces

Discovering kinship relations from face images in the wild has become an interesting and important problem in multimedia and computer vision. Despite the rapid advances in face analysis in unconstrained environment, kinship verification still remains a challenging problem as the subtle kinship relation is difficult to discover and changes in pose and lighting condition further complicate this task. In this paper, we propose a kinship verification approach based on multi-linear coherent space learning. Local image patches at different scales are independently projected into their corresponding coherent spaces learned by robust canonical correlation analysis such that patch pairs with kinship relations have improved correlation. In addition, most discriminative patches for verification are selected via constrained linear programming. Experimental results on two widely used kinship verification datasets show that the proposed method can effectively identify different kinship relations in image pairs. Compared to state-of-the-art techniques, the proposed method achieves very competitive performance with the use of simple feature descriptors.