Social relationships classification using social contextual features and SVDD-based metric learning.

Family relationship is an important concern in image-based social relationships recognition, and there are very limited attempts to tackle diverse social relationships in the literature. In this paper, we propose the problem of social relationships classification in which we aim to model three types of social relationships( e.g., family, colleagues and friends) in the images. To this end, we introduce two types of social contextual features to capture detailed information( e.g., geometry or appearance) in images. Moreover, we present a new Support Vector Data Description-based metric learning( SML) method for social relationships classification. Motivated by the fact that the images are unavoidably degraded by noise due to some variation factors such as illumination and pose, we aim to learn a robust distance metric to suppress noise and model the spatial structure among multiple entities, such that more discriminative information can be exploited for classification. We also extend our method to multiview version-MSML, which helps to exploit multiple features to improve the social relationships classification performance. Extensive experiments on our newly released social relationships database demonstrate the feasibility and effectiveness of our proposed methods.