ABSTRACT
In this chapter we discuss the process of extracting face-specific descriptors from a range data description of a head. The situation is the same as that described in the last section of the previous chapter. The data was acquired by a Cyberware scanner in cylindrical coodinates (r, 0, y) based on a vertical y-axis through the center of the head and consists in measurements r = f(6i,yj),l < i < 512,1 < j < 256 including the whole head. We compute the principal curvatures, parabolic and ridge curves as in the previous chapter. The goal now is to extract important scalar features for the purpose of discriminating individual faces. In particular, the following features are extracted:
• nose bridge (nasion) • nose base (base of septum) • nose ridge
• eye corner cavities: inner and outer
• convex center of the eye (eyeball/lid region) • eye socket boundary (surrounding the convex center of the eye) • boundary surrounding nose
• face region
ships among depth and curvature features. As such, the feature extraction process is structured as a constrained search in the crowded space of low-level features. As each constraint is enforced, the number of potential locations for the feature
is reduced. The constraints are designed to reduce the search to a single definition of the feature. At that point, the feature's relationship to other previously extracted features is examined as a verification of its detection. If error conditions are detected during the search or validation, no feature value is recorded. The constraints for one feature are often functions of other related features' values or estimates of these values if a feature is not available. Thus, as the feature detection progresses, these constraints become more well-defined. In some cases, particularly if the previous calculation of a feature produced errors or if the feature value is particularly sensitive in other processes (feature detection or recognition), we extract a feature's description again when more accurate constraints are available. The search process considers constraints on:
• absolute extent of a region on the surface;
• distance of a region or point from the symmetry plane;
• symmetrical pairing of features;
• proximity of a point or region to a target on the surface;
• amount of protrusion of a region from the surrounding surface;
• local configuration of elliptic blue (maximum) ridge lines and elliptic red (minimum) ridge lines (Section 6.4.1 and Section 6.4.2). Here we take the face normal to point inwards. The elliptic red ridge lines will also be called valley lines.
