Kinship similarity for open sets

The aim of image-based kinship recognition methods is to determine the genetic relationship between people from their face images. Kinship recognition and related methods have many applications in computer vision. Most of the methods are very meaningful but are focused on closed sets. However, for real-life scenarios, kinship recognition is an open set problem.
In contrast to previous methods, this paper considers kinship recognition for open sets. Further, the aim is to determine family relationships and their corresponding degrees of kinship. Our method is pairwise-based and is able to exploit mutual information from positive pairs.
Large scale experiments and ablation studies show that our method (1) reaches SOTA performance on the FIW dataset, (2) is able to properly separate kinship categories using pairwise similarity and (3) generates uniform similarity distributions.