Quantitative Facial Landmark Analysis for Gender Scoring

We annotated 23 biologically significant landmarks [28] on each image as shown in Figure 2. The motivation for using these landmarks comes from the fact that they represent the sexual dimorphism of the face [29] (link). These landmarks and Euclidean distances measured from them are used to measure a quantitative dimension for the morphological deviation from the normal face [28] , to delineate syndromes [30] (link) and to measure objective masculinity/femininity [21] (link). We have selected the facial landmarks that relate to the bony structure of the face which is effected by the ratio of testosterone to estrogen (oestrogen) during adolescence [31] (link). It is believed that facial masculinity is associated with levels of circulating testosterone in men [19] (link). Hence it is intuitive to use features extracted from these bony landmarks for facial gender scoring.
The pose of each 3D face is corrected to a canonical form based on four landmarks (Ex(L), Ex(R), N and Prn). This step is required to eliminate any error due to pose in the extraction of geodesic distances which will be discussed in detail in the Study 2 of the Experiments Section. Holes are filled and noise removed by re-sampling the 3D face on a uniform grid using the gridfit [32] algorithm. Since some portions of the face are expected to be self occluded (e.g. region around Ac) when re-sampled on a grid, we bisect the 3D face along the vertical axis at the nose tip and rotate each half by before re-sampling to mitigate this problem. Besides hole filling, another advantage of bisecting and rotating the halves before re-sampling is that the resulting 3D face has a more uniform sampling in the 3D space. The processed halves are then rotated back and stitched seamlessly to form a single mesh. Figure 3 shows the different preprocessing steps.