Adversarial Similarity Metric Learning for Kinship Verification

Given a pair of facial images, it is an interesting yet challenging problem to determine if there is a kin relation between them. Recent research on that topic has made encouraging progress by learning a kin similarity metric from kinship data. However, most of the existing metric learning algorithms cannot handle hard samples very well, i.e., some ambiguous test pairs cannot be well classified due to some compounding factors, such as the large age gap or gender difference between the parents and children. To address this, we propose an Adversarial Similarity Metric Learning (ASML) method in this paper. More specifically, ASML consists of two adversarial phases: confusion and discrimination. In confusion phase, ambiguous adversarial pairs are automatically generated to challenge the learned similarity metric; while in discrimination phase, the learned metric tries its best to adjust itself to distinguish both the original pairs and the generated adversarial pairs. Consequently, a robust and discriminative similarity metric can be learned by iteratively performing the two adversarial phases. Experiments on the two widely used kinship datasets demonstrate the efficacy of our proposed ASML method in comparison with the state-of-the-art metric learning solutions to kinship verification.