Normalization And Shape-Recognition Of 3-Dimensional Objects By 3D Moments

Abstract

In this paper we are primarily concerned with the recognition of three-dimensional (3D) objects on the basis of the geometry of their physical surface. The 3D moments are defined on the object's surface with two main aims: (1) to derive a normalized version irrespective of position, size, and orientation for each imaged object; and (2) to establish global descriptors, extracted from normalized shapes, in order to make up a representative feature vector to be used in recognition tasks. The source data are multiple orthographic views of 3D objects with known viewpoint specifications. A volume intersection procedure is used in the recovery of 3D object surface from two-dimensional (2D) data. The problem of normalization to achieve recognition of 3D objects is dealt with and a heuristic solution is proposed to lessen the inherent ambiguity of the principal axes method for object orientation. Experimental results are described with ten object categories, showing excellent percentage classification success rates by using only a small number of normalized moments as elements of the feature vector.