Curvature based localization of nose tip point for processing 3D-face from range images

Unconstrained acquisition of data from arbitrary subjects results in facial scans with significant pose variations. The challenges in 3D face recognition are into two main stages, namely preprocessing range scans for detection of fiducial detection while identifying/filling missing parts due to occlusions along with outlier noise reduction and during post-processing where actual match is done with stored models. In this work, an algorithm using HK curvature for localization of nose tip fiducial point on 3D-face image is proposed at preprocessing stage. Curvature is evaluated on 3D data following the normalization step. HK curvature classification results potential region segmentation on face and operated further with morphological enhancements. Four types of curvatures- elliptical convex, elliptical concave, hyperbolic convex and hyperbolic concave enhanced curvature profiles are being processed separately. Coarse-to-fine scale space using integral images technique is applied on the curvature images. Localization is boosted using a heuristic driven bag of templates rule. The proposed technique achieved up to 90% accurate nose-tip localization on Gavabdb and FRAV3D face database.