Overview

Contrary to other data modalities, processing 3D point clouds poses several significant challenges. Point cloud data generated by inexpensive sensors suffers from the presence of artifacts, non-uniform noise, and variation in density. Established 2D image processing techniques are not directly applicable to 3D point clouds. The main reasons for this are the following: (i) difference in data representation - an image is organized as a matrix while a 3D point cloud is an unorganized and irregularly distributed set of points; (ii) difference in information presentation - an image contains ambiguous spatial information and abundant spectral information, while a 3D point cloud contains explicit spatial information and possibly no spectral information; (iii) difference in spatial neighborhood - an image is arranged in a grid-like pattern thus allowing the neighborhood of a pixel to be easily determined, conversely the neighborhood of 3D point cloud must be determined by a nearest neighbors search.

Contributions

• We created a new global shape descriptor based on computing the topological persistence of the zeroth and first homology groups of a 3D point cloud
• We developed a novel topology-based information gain metric for directing the next best view of a noisy 3D sensor

Publications

1. W.J. Beksi and N. Papanikolopoulos, "A Topology-Based Descriptor for 3D Point Cloud Modeling: Theory and Experiments," Image and Vision Computing, 2019.
   Paper  •   Citation
2. C. Collander, W.J. Beksi, and M. Huber, "Learning the Next Best View for 3D Point Clouds via Topological Features," International Conference on Robotics and Automation (ICRA), 2021.
   Paper  •   Preprint  •   Source Code  •   Citation