3D model retrieval based on linear prediction coding in cylindrical and spherical projections using SVM-OSS

Abstract

This paper presents a 3D model descriptor based on the linear prediction coding (LPC) coefficients to retrieve 3D objects. In this method, the 3D object is projected on the lateral surface of a cylinder parallel to one of its principal axes and centered at the centroid of the object. To improve efficiency, cylindrical projection is performed on three cylinders parallel to principal axes. Besides, the object is projected on a sphere centered at the centroid of the 3D object. The surface of object is converted to two-dimensional shapes as a result of performing the projections which preserves the geometric features of the object. Then, LPC coefficients are extracted from the two-dimensional projected shapes. These coefficients estimate the parameters of correlated signal, efficiently. Since the cylindrical and spherical projections of the 3D model are correlated surfaces, LPC coefficients describe the surfaces as well. The rotation normalization is performed employing the principal component analysis. Furthermore, the retrieval performance is enhanced employing SVM-OSS similarity measure which efficiently compares two model feature vectors. We implement an experimental comparison on PSB database of 3D models in order to demonstrate the performance of the proposed descriptor. Experimental results show the effectiveness of the proposed descriptor compared to current methods.