Efficient object categorization with the surface-approximation polynomials descriptor

Perception of object categories is a key functionality towards more versatile autonomous robots. Object categorization enables robots to understand their environments even if certain instances of objects have never been seen before. In this paper we present the novel descriptor Surface-Approximation Polynomials (SAP) that directly computes a global description on point cloud surfaces of objects based on polynomial approximations of surface cuts. This descriptor is directly applicable to point clouds captured with time-of-flight or other depth sensors without any data preprocessing or normal computation. Hence, it is generated very fast. Together with a preceding pose normalization, SAP is invariant to scale and partially invariant to rotations. We demonstrate experiments in which SAP categorizes 78% of test objects correctly while needing only 57ms for the computation. This way SAP is superior to GFPFH, GRSD and VFH according to both criteria.