Analysis of a multi-ring technique for 3D pattern recognition

Pattern characterization of geometric features is of great interest in the cultural heritage domain, as the style of the some parts such beards, hairs, helms or cuirass decorations would support and enhance the classification and/or cataloguing of the archaeological artefacts and fragments. In this context, a pattern is a geometric or colour feature which is repeated over a surface, either on its entirety or on a part (for instance, engravings and chisellings, ornamental decorations, etc.). In the literature, methods that face this problem either project the surface into an image and then adopt a image pattern recognition techniques or try to extend in 3D an existing image technique to surfaces. In this paper, we analyse the performance of the technique called mesh Local Binary Pattern (meshLBP) that works on triangulation meshes, considering implementation defined in [WTBdB15, WBdB15] and freely available1. Tests are performed on meshes of the repository at STARC - The Cyprus Institute [STA] that correspond to laser scans of earthenware fragments retrieved from the Salamina Island. Fragments are represented with triangulations, not necessarily uniformly sampled. In this paper we analyse two kind of patterns represented by geometric and colorimetric variations on the surface of the fragment. In addition, we discuss a sub-sample strategy to keep the effectiveness of the meshLBP descriptor and the dependence of the operator on parameters such as the surface representation and the number of rings used for its multi-scale evaluation. Finally, the main limitations of the method are discussed and possible improvements and future developments are outlined.