Figure 1.

Pipeline to measure facial femininity in humans and mandrills

DCNN model (VGGFace) was first retrained with FaceScrub database³⁷ for humans and the Mandrillus Face Database⁴⁰ for mandrills to simultaneously perform identity verification and sex classification (i.e., multi-task learning). Portrait pictures of the studied individuals were then fed to the newly trained DCNN, and their features extracted. The 128-d penultimate layer of the DCNN was used as a face space. Images of CFD database (for analyses on humans) and Mandrillus Face Database restricted to adults (for analyses on mandrills) were projected into their respective face space. Retrieved features were reduced further to two dimensions using a PCA.The third box (from the left) represents the position of different pictures of different individuals in the PC1-PC2 plane (for clarity, only twenty individuals are presented; one shape of a given color corresponds to one individual).While PC1 and PC2 together allowed us to cluster portraits collected on the same individual, PC1 alone discriminated the two sexes. We thus defined the female centroid, cf, and the male centroid, cm, as the mean PC1 score of all female and male pictures, respectively. We then calculated two measures of femininity for a given portrait picture i: (1) dcf, the distance between PC1_i and cf, which describes the averageness of a face within the female faces’ category, and (2) dcm, the distance between PC1_i and cm, which describes the dimorphism of females’ faces compared to males’ faces. In this depiction, a highly feminine face i is described by a high value of dcm_i and a small value of dcf.