Nose tip localization on 2.5D facial models using differential geometry based point signatures and SVM classifier

Nose tip localization is often the basic step for 2.5D face registration and further 3D face processing and as such appears as a side problem of most research works on 2.5D or 3D face recognition. In this paper, we propose to carry out a comprehensive study of four popular rotation invariant differential geometric properties, namely Mean and Gaussian curvature, Shape Index and Curvedness, for the purpose of nose tip localization. For each 2.5D facial model, the set of nose tip candidates is first automatically selected from a shape classification thanks to a priori knowledge of a nose region. A SVM classifier trained on a subset of the data set using the previous four curvature descriptors alone or in combination is then invoked to select the true nose tip from the candidate set. We report extensive experimental results cross-validated in terms of True Acceptance Rate (TAR) and False Acceptance Rate (FAR) in comparison with manually labeled nose tip as the ground truth. A 99.9% Nose Tip TAR with 6.71% FAR is achieved on the FRGC v2.0 dataset when Mean curvature and Shape Index along with Curvedness are used as the input to the SVM.