3D SHAPE COMPLETION WITH DEEP NEURAL NETWORKS

Abstract

This thesis studies deep learning-based methods to improve perception ability in the 3D domain. Among different kinds of 3D data, point clouds are widely used in 3D related tasks. However, the raw data of point clouds are often sparse, irregular and incomplete, which lead to difficulties on the down-streaming tasks, such as object classification, detection and segmentation. To alleviate the above problems, this thesis introduces three methods to solve the shape completion problem in 3D point clouds. For 3D point cloud completion, we aim to generate dense and complete point clouds from partial and sparse inputs. We propose to improve the object completion quality by leveraging the prior knowledge from both the partial and complete scans. Moreover, a self-supervised method is proposed. Our approaches are intensively evaluated in various experimental settings and show superior performances both qualitatively and quantitatively.