Partial 3D Object Retrieval combining Local Shape Descriptors with Global Fisher Vectors

Abstract

This work introduces a partial 3D object retrieval method, applicable on both meshes and point clouds, which is based on a hybrid shape matching scheme combining local shape descriptors with global Fisher vectors. The differential fast point feature histogram (dFPFH) is defined so as to extend the well-known FPFH descriptor in order to capture local geometry transitions. Local shape similarity is quantified by averaging the minimum weighted distances associated with pairs of dFPFH values calculated on the partial query and the target object. Global shape similarity is derived by means of a weighted distance of Fisher vectors. Local and global distances are derived for multiple scales and are being combined to obtain a ranked list of the most similar complete 3D objects. Experiments on the large-scale benchmark dataset for partial object retrieval of the shape retrieval contest (SHREC) 2013, as well as on the publicly available Hampson pottery dataset, support improved performance of the proposed method against seven recently evaluated retrieval methods.