3D Point Cloud Restoration via Deep Learning: A Comprehensive Survey

Abstract

<p indent=0mm>3D point cloud is one of the most commonly used 3D scene/object representation methods. According to the different emphases of point cloud restoration, 3D point cloud restoration technologies based on deep learning are divided into three classes: dense reconstruction, complete reconstruction and denoising reconstruction. Typical restoration models and key techniques, such as feature coding, feature extension, and loss function design, are analyzed. Commonly used network modules, point cloud data sets, and evaluation criteria are summarized. Finally, the relationship between the three kinds of point cloud restoration technologies is discussed, and the challenges and future development trends of point cloud restoration technologies are explored from five aspects: rotation invariant feature extraction, detail information repair, topological relationship preservation, geometric algorithm application, and multimodal data fusion.