Can faces verify blood-relations?

Humans can verify unknown parent-offspring and sibling pairs over unrelated subject pairs. A computational scheme to accomplish the task robustly, in the presence of challenges due to gender and age gap between related-pairs, finds many applications such as matching orphaned/lost children, identifying relatives from a photo collection. We propose one of the first computational schemes to verify sibling pairs along with parent-child relation. Towards the same, we present a novel ensemble metric learning scheme that combines the advantages of task-specific learning, adaptive prototype and feature selection and `late fusion´. We demonstrate the robustness of the scheme on a very large scale, real-world dataset. Specifically, we show that the gender difference among related-pairs leads to lower performance of traditional verification and metric learning algorithms. Through various experiments, we quantitatively study the robustness of the proposed scheme in the general and specific case of different gender related-pairs, achieving up to 80%, 75% accuracy for the parent-child and siblings relations respectively.