A Similarity Model for 3D Objects Based on Stable Sub-clouds

Abstract

We present an idea of a novel similarity model for objects represented by 3D point clouds that were generated by scans of real-world objects. Various existing approaches find descriptive points on the object surface or extract features of groups of points. However, 3D object scans when conducted outside a lab environment often suffer from imprecisions and noise artifacts, which many existing approaches do not handle well. To better tolerate these imperfections, our model extracts stable sub-clouds from the input point cloud, which represent classes of adjacent sub-clouds sharing similar features. We demonstrate experimentally that features generated from these sub-clouds can be used to establish a measure of similarity between objects. We show preliminary results of an application of this technique to point clouds of models scanned from real-world objects and demonstrate that this technique has good potential to deal with imperfect data by showing how the computed distance relates to degrees of modification of the data. Our technique extracts features from particularly resilient portions of the object and is thus better able to accommodate deficiencies in the input data.