Point-Triplet Spin-Images for Landmark Localisation in 3D Face Data

This paper introduces and evaluates our point-triplet spin-image descriptor, a novel descriptor that requires three vertices to be computed. This descriptor is able to encode surface information, within a spherical neighbourhood with radius r defined from a triplet's baricenter, into a surface signature. We believe that this new descriptor could be useful within a number of graph based retrieval applications; however, here we evaluate its performance within 3D face processing in the first instance. In doing so, this descriptor is embedded into a system designed to simultaneously localise the nose-tip and the two inner-eye corners of a human face. First, candidate triplets are gathered using the structured graph matching approach “relaxation by elimination” with a basic graph of three vertices and three arcs. Next, these candidate landmark-triplets are evaluated as in a binary decision problem. Hence, a point-triplet spin-image feature for each candidate landmark-triplet is computed and evaluated according to its Mahalanobis distance. This investigation includes two state of the art datasets, the Face Recognition Grand Challenge (FRGC) and CurtinFaces, as well as a performance comparison between this point-triplet spin-image and another point-triplet descriptor, named weighted-interpolated depth map which give us promising results and encourages our face processing research.