UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Point cloud classification by nearest samples of random rays Liu, Liangchen

Abstract

Point clouds serve as a common type of data representation for general geometric objects and abstract data structures in arbitrary dimensions. With the accessibility of reliable and efficient 3D scanning technology as well as the pervasiveness of data science, point clouds widely appear in various research areas and applications, especially in geometric processing, computer vision, and robotics, making robust point cloud processing algorithms an urgent need. However, due to the inherently irregular structure of point clouds, the remarkable success of many traditional algorithms cannot be duplicated directly on point cloud processing. Besides, semantic understanding of underlying shapes, which is essential in many applications in computer vision, requires interpretation of refined information, yet how to properly extract useful geometric information from point clouds is still an open problem in this field. To this end, in this thesis, we focus on point cloud classification, one of the fundamental point cloud processing tasks, and present a novel framework, RaySense$+$RayNN, to conquer this challenge. First, we use RaySense, an innovative subsampling strategy, to compute the signature of a target point cloud by finding the nearest neighbors of points on a set of randomly generated rays. Using rays, we incorporate computational structure into the point cloud thereby different strategies and algorithms can be implemented. We then propose a convolutional neural network, RayNN, that operates on the RaySense signature to perform point cloud classification. By analysis and experiments, we reveal the RaySense signature is not merely a subsample of the original point clouds, it also preserves permutation invariance of the input points and extracts certain statistical as well as geometric information, independent of the choice of ray set. We show that even with the simple architecture, our framework can achieve comparable performance against the well-tuned state-of-the-art algorithms on the benchmark dataset. Furthermore, we explore the robustness of our framework by exposing it to a variety of data corruptions, indicating its capacity for practical applications. Lastly, we conclude with a discussion on potential applications and directions for future research.

Item Media

Item Citations and Data

Rights

Attribution-NonCommercial-NoDerivatives 4.0 International