Evaluation of Local 3-D Point Cloud Descriptors in Terms of Suitability for Object Classification

This paper investigates existing methods for local 3-D feature description with special regards to their suitability for object classification based on 3-D point cloud data. We choose five approved descriptors, namely Spin Images, Point Feature Histogram, Fast Point Feature Histogram, Signature of Histograms of Orientations, and Unique Shape Context and evaluate them with a commonly used classification pipeline on a large scale 3-D object dataset comprising more than 200000 different point clouds. Of particular interest are the details of the choice of all parameters associated with the classification pipeline. The point clouds are classified by using support vector machines. Fast Point Feature Histogram proves to be the best descriptor for the method of object classification used in this evaluation.