Fig. 4. Pipeline for measuring facial resemblance.

The pipeline contains two main steps. During the face identification step, a deep neural network previously trained for human face recognition (VGG-Face) is retrained to identify mandrill faces. The newly trained network is then used in a face verification task, first to learn a distance metric using a support vector machine (SVM) trained to detect whether two faces represented by their feature activation vectors (i.e., coordinates in the DFS) represent the same individual or not, and then to compute the resemblance (i.e., the distance in the DFS) between pairs of portrait images for the studied population.