Gender-based Feature Disentangling for Kinship Verification

Kinship verification can benefit a wide variety of applications, e.g., exploring social relations, finding the lost children and old people, constructing a family tree, and so on. Previous researches have made promising results in this research, but the gender discrepancy between parent and child is generally neglected. For example, father and daughter, or mother and son, may have different facial features due to gender differences. In view of this, we propose a gender-invariant kinship verification model where the facial feature is divided into two components. i.e., gender-dependent feature and identity-dependent feature. The learning of gender-dependent feature is supervised by the gender prediction task. This identity-dependent feature is required to be uncorrelated to the gender-dependent feature and preserve information that is useful for kinship verification. We factorize facial features through a Residual Factorization Module (RFM) and reduce the correlation between two components through the Decorrelated Adversarial Learning (DAL). The whole network is trained in an end-to-end and multi-task manner. Experimental results on the popular benchmark KinFaceW-II demonstrate that our gender invariant features can effectively reduce the effects of gender differences and show excellent generalization ability on different kinship relations.