Deep learning-based techniques for 3D point cloud analysis

Abstract

This thesis presents two innovative works in the field of point cloud processing: ODFNet and ALReg. The ODFNet work proposes a new method for the classification and segmentation of point clouds, while ALReg aims to facilitate the training of neural networks for point cloud registration using active learning. Convolutional neural networks (CNNs) have been widely used for visual tasks such as object categorization, object detection, and semantic segmentation. However, the application of CNNs to point clouds is a relatively recent development. Notably, there has been no previous approach that specifically utilizes point density information. To enhance the representational power of local features, we propose leveraging the distribution of point orientations within a neighborhood relative to a reference point. This led us to introduce the concept of point Orientation Distribution Functions (ODFs). In order to compute Orientation Distribution Functions (ODFs), our approach involves dividing the spherical region around each point into a set of cones aligned with predefined orientations. Within each cone, we calculate the density of points, resulting in the ODFs for that particular point. These ODF patterns provide valuable information about the spatial structure and orientation characteristics of the objects within the point cloud. The ODFs allow us to summarize the local neighborhood structure of the point cloud in a concise manner, which is beneficial for our point cloud analysis network model design. By incorporating the ODFs, we can effectively capture and utilize the significant local geometric information present in the point cloud, leading to improved performance and accuracy in our analysis tasks. Additionally, we introduce the ODFNet, a neural network model specifically designed for point cloud analysis tasks. The ODFNet utilizes the ODFBlock and ODFs to effectively incorporate the directional information captured by the ODFs. This integration enhances the performance of various tasks, such as classification, segmentation, and other related applications, by leveraging the rich geometric information provided by the ODFs. The experimental results confirm that ODFNet achieves state-of-the-art performance in both classification and segmentation tasks. Point cloud registration, which involves aligning multiple point clouds by calculating the rigid transformation between them, is a crucial task in computer vision. While there have been significant advancements in point cloud registration using various methods, most of them rely on training the network with the entire shape for registration. However, an alternative approach that has not been explored in the literature is active learning, which could also be employed to address this problem. Thus, in this thesis, we also introduce ALReg, an active learning approach aimed at improving the efficiency of point cloud registration by selectively utilizing informative regions to reduce training time. To achieve this, we modify the baseline registration networks by incorporating Monte Carlo DropOut (MCDO) to efficiently calculate uncertainty. Our main objective is to demonstrate that similar accuracy scores can be achieved by using fewer point clouds or parts of point clouds in the training phase of any point cloud registration network, provided that the selection of these training samples/parts is done effectively. By leveraging ALReg, our goal is to incorporate the most effective point cloud parts into the training procedure, thereby improving the efficiency and effectiveness of the registration network.