Heterogeneous Similarity Learning for More Practical Kinship Verification

Kinship verification via facial images is a relatively new and challenging problem in computer vision. Prior studies in the literature have focused solely on gender-fixed kin relation, i.e., on the question of whether one gender-fixed kin relationship between given subjects can be established. In practice, however, large scale gender annotation is time-consuming and expensive. Instead, we propose in this paper to learn and predict with gender-unknown kin relations. To address this, we present a novel heterogeneous similarity learning (HSL) method. Motivated by the fact that different kinship relations may not only share some common genetic characteristics but also have its own inherited traits from parents to offspring, we aim to learn a similarity function under which the commonality among different kinship relations are captured and the geometry of each relation is preserved, simultaneously. We further derive a multi-view HSL method by optimal fusion of the similarity models from multiple feature representations, such that the complementary knowledge in multi-view kin data can be leveraged to obtain refined information. Experimental results demonstrate the effectiveness of our proposed methods.