Point-pair descriptors for 3D facial landmark localisation

Abstract

Our pose-invariant point-pair descriptors, which encode 3D shape between a pair of 3D points are described and evaluated. Two variants of descriptor are introduced, the first is the point-pair spin image, which is related to the classical spin image of Johnson and Hebert, and the second is derived from an implicit radial basis function (RBF) model of the facial surface. We call this a cylindrically sampled RBF (CSR) shape histogram. These descriptors can effectively encode edges in graph based representations of 3D shapes. Thus, they are useful in a wide range of 3D graph-based retrieval applications. Here we show how the descriptors are able to identify the nose-tip and the eye-corner of a human face simultaneously in six promising landmark localisation systems. We evaluate our approaches by computing root mean square errors of estimated landmark locations against our ground truth landmark localisations within the 3D Face Recognition Grand Challenge database.