Kinship Verification Using Hierarchical Structures and Extended Contrastive Learning
Eran Dahan, Yosi Keller
Abstract
In this work, we aim to improve kinship verification performance by optimizing embedding representations tailored to each kinship relation type. We concentrate on two relationship categories: samegeneration (e.g., Brothers, Sisters, Siblings) and mixed-generation (e.g., Father-Daughter, Mother-Son). For mixed-generation relationships, we develop a sophisticated contrastive learning framework that takes advantage of the hierarchical structure within a family, such as refining the kinship relation embedding for Mother-Daughter as an extension to the Sisters relationship. For the types of same-generation relationships, we propose a tailored contrastive learning scheme for each specific kinship relationship. Further, we developed a unique sampling method for our scheme which helps to reduce the overfitting of the kinship verification task. Overall, our method achieves state-of-the-art performance on the FIW dataset, outperforming previous benchmarks by a substantial margin.