Which 3D geometric facial features give up your identity?

The 3D face recognition literature has many papers that represent facial shapes as collections of curves of different kinds (level-curves, iso-level curves, radial curves, profiles, geodesic polarization, iso-depth lines, iso-stripes, etc.). In contrast with the holistic approaches, the approaches that match faces based on whole surfaces, the curve-based parametrization allows local analysis of facial shapes. This, in turn, facilitates handling of pose variations (probe image may correspond to a part of the face) or missing data (probe image is altered by occlusions. An important question is: Does the use of full set of curves leads to better performances? Among all facial curves, are there ones that are more relevant than others for the recognition task? We explicitly address these questions in this paper. We represent facial surfaces by collections of radial curves and iso-level curves, such that shapes of corresponding curves are compared using a Riemmannian framework, select the most discriminative curves (geometric features) using boosting. The experiment involving FRGCv2 dataset demonstrates the effectiveness of this feature selection by achieving 98.02% as rank-1 recognition rate. This selection also results in a more compact signature which significantly reduces the computational cost and the storage requirements for the face recognition system.