Volume Net: Flexible Model for Shape Classes

This paper proposes a novel approach for modelling 3D shape classes. The approach is based on approximating 3D shapes by polyhedrons created by unions of intersections of convex sets and complements of convex sets. Models of shape classes are created from training samples in the form of 3D triangular meshes. These models can be used to classify objects in depth images in one of previously learned object classes and to provide information about their pose and shape parameters. This ability of the proposed approach is its advantage over the most of the object classification methods, since it allows adaptation of a robot action defined for a referent model of an object class to the other instances of the same class. Application of the proposed approach for object classification is tested using a challenging 3DNet dataset. The obtained classification results are competitive to four well-known 3D shape descriptors.