Research on deep learning algorithms for point cloud upsampling

Abstract

The integration of 3D vision technology has profoundly impacted various aspects of our lives, with point clouds emerging as the predominant geometric representation in the field due to their simplicity, ease of processing, and low acquisition costs. However, limitations in 3D scanning hardware often result in sparse, noisy, and unevenly distributed point clouds, which impede the effectiveness of subsequent 3D vision tasks. To obtain high-quality and precise point clouds, researchers have introduced upsampling algorithms that enhance point cloud representativeness through densification. In recent years, deep learning has supplanted traditional algorithms that rely on a priori data, owing to its robust learning representation capabilities and data-driven nature. Consequently, deep learning algorithms for point cloud upsampling have become a highly valuable research direction within the 3D vision domain. This paper delves into recent advancements in deep learning algorithms for point cloud upsampling by first defining the problem and analyzing the technical challenges from six perspectives. It then provides an overview of the relevant benchmarks in point cloud upsampling and traces their developmental history. The paper further examines the technical issues in the three critical steps (point cloud local feature extraction, feature space expansion, and coordinate reconstruction) of the overall deep learning algorithm for point cloud upsampling. It categorizes the prevalent technical solutions, discussing their merits and drawbacks, and offers optimization strategies. Lastly, the paper presents a forward-looking perspective on potential research directions in this field from four angles.