CSIOR: An Ordered Resampling Approach for Relief Pattern Analysis using Image-like Operators on the Mesh Manifold

  • Claudio Tortorici

Student thesis: Doctoral Thesis

Abstract

In the last few years, an increasing number of capturing devices appeared on the market which are capable of acquiring, either statically or dynamically, real objects geometry. Typically, such 3D scanners produce clouds of irregularly distributed 3D points that can be connected in order to obtain a polygonal approximation (i.e.,mesh manifold) of the target surface. In image processing, 2D analysis often adopt local descriptors, see Local Binary Pattern and convolution-based approaches, which relies on the ordered structure of the image grid. Our aim is to address the problem of efficiently handling mesh manifolds, focusing on 3D Face Recognition and Relief Pattern analysis. While face scans are renowned, relief patterns are a quite new area of investigation. Such patterns are corrugation of the object surface, that do not alter its overall shape: therefore, can be considered as the 3D counterpart of the texture concept in 2D images. In this thesis, we developed an LBP-based approach which is capable of per- forming face recognition on mesh manifold directly; this lead to the possibility of fusing different geometric features and of the face photometric information at the mesh level, improving the recognition, being independent to observation point and lighting which affect images. Then, with the aim of retracing the success of 2D LBP on the 3D space, we propose a framework capable of casting a wide range of LBP variants, thus expanding the range of 3D descriptors which operate directly on mesh manifold. Also, we propose an efficient approach to perform convolution, which allow the discrete and continuous filtering of mesh surfaces. These approaches have been widely tested showing remarkable results in Relief Pattern Retrieval and Classification tasks. Finally, we propose a novel algorithm for mesh regularization, dubbed CSIOR, which is distinguished by its intrinsic capability to produce a globally ordered triangular mesh made by quasi-equilateral facets. This tool revealed to be useful for the above-mentioned tasks as it provides a regular mesh tessellation while preserving the shape of the object and its surface geometric characteristics.
Date of AwardDec 2019
Original languageAmerican English
SupervisorNaoufel Werghi (Supervisor)

Keywords

  • CSIOR; Mesh Manifold; Relief Patterns; 2D LBP; 2D Images.

Cite this

'