Exploring Automatic Face Recognition on Match Performance and Gender Bias for Children

In this work we update the body of knowledge on the performance of child face recognition against a set of commercial-off-the-shelf (COTS) algorithms as well as a set of government sponsored algorithms. In particular, this work examines performance of multiple deep learning face recognition systems (8 distinct solutions) establishing a performance base line for a publicly available child dataset. Furthermore, we examine the phenomenon of gender bias as a function of match performance across the eight (8) systems. This work highlights the continued challenge that exists for child face recognition as a function of aging. Rank-1 accuracy ranges from 0.44 to 0.78 with an average accuracy of 0.63 on a dataset of 745 unique subjects (7,990 total images). Furthermore, when we introduce a distractor set of approximately 10; 000 child faces the rank-1 accuracy decreases across all systems on an average of 10 points. Additionally, the phenomenon of gender bias is exhibited across all systems, although the developers of the face recognition systems claim a near balance of genders was used in the development. The question of gender disparity is elusive, and although co-factors such as makeup, expression, and hair were not explicitly controlled, the dataset does not contain substantial differences across the genders. This work contributes to the body of knowledge in multiple categories, 1. child face recognition, 2. gender bias for face recognition and the notion that females as a sub-population may exhibit Lamb characteristics according to Doddington's Biometric Zoo, and 3. a dataset for child face recognition.