Mining Contrastive Loss for Kinship Verification

Facial kinship verification is a task that aims to rec-ognize biological relationships between individuals based on facial images. Kinship verification is challenging because it requires identifying subtle similarities between relatives. The supervised contrastive loss applied to kinship verification generates a large number of redundant pair-wise samples. As an improvement, we propose a mining contrastive (MC) loss to enhance the discriminative ability of contrastive loss through a hard sample mining strategy which emphasizes the balance between positive and negative samples, improving overall verification accuracy. Compared with supervised contrastive loss, our proposed MC loss achieves better performance on FIW dataset.