The fast growth of 3D content available on Internet leads to the need of developing new efficient analysis tools for organizing and making use of 3D data. Most of these 3D models exhibit some symmetry that can be useful for their identification. However, the symmetry are often approximate and local and therefore hardly detectable. In this thesis, we suggest an original multi scale symmetry detection method for 3D surfaces made of polygonal meshes. This method will then be applied to the calculation of a new 3D descriptor for indexing and classification purpose.

In the first part of this thesis, we introduce some multi scale feature metrics for surface points, in order to match them efficiently and thus make a more reliable symmetry detection algorithm. The resulting probabilistic distribution is then analyzed with a regular sampling of the symmetry plane space through a 3D Hough Transform. Experimental results show a significant decrease of processing load for mirror symmetry detection. In the last part of this thesis, we will investigate the definition of a new model descriptor based on its global and local symmetries and we will study how it can improve both indexing and feature point extraction performance.