Augmenting Automated Kinship Verification with Targeted Human Input.

Kinship verification is the problem whereby a third party determines whether two people are related. Despite previous research in Psychology and Machine Vision, the factors affecting a person's verification ability are poorly understood. Through an online crowdsourcing study, we investigate the impact of gender, race and medium type (image vs video) on kinship verification - taking into account the demographics of both raters and ratees. A total of 325 workers completed over 50,000 kinship verification tasks consisting of pairs of faces shown in images and videos from three widely used datasets. Our results identify an own-race bias and a higher verification accuracy for same-gender image pairs than opposite-gender image pairs. Our results demonstrate that humans can still outperform current state-of-the-art automated unsupervised approaches. Furthermore, we show that humans perform better when presented with videos instead of still images. Our findings contribute to the design of future human-in-the-loop kinship verification tasks, including time-critical use cases such as identifying missing persons.