Research and application on deep learning-based point cloud completion for marine structures with point coordinate fusion and coordinate-supervised point cloud generator

Abstract

The problem of missing point clouds is prevalent in the actual point clouds of Marine Structures (MS) obtained based on three-dimensional laser scanning technology. To achieve the completion tasks for MS, this paper proposes a deep learning network, MS-PCN, and builds a point cloud completion dataset, MS-dataset. MS-PCN employs both point coordinate fusion module and coordinate-supervised point cloud generator to improve the accuracy of point cloud completion for MS. Extensive experiments conducted on MS-dataset and public dataset ShapeNet-55 demonstrate the effectiveness of MS-PCN in point cloud completion within scenarios featuring MS as well as its generalizability in other scenarios. MS-PCN achieved a Chamfer Distance (CD) of 0.31 and an F-score of 0.58 on MS-dataset and a CD of 0.70 and an F-score of 0.505 on ShapeNet-55 dataset. Furthermore, point cloud completion could serve as a valuable precursor to the surface reconstruction of MS, improving its reconstruction accuracy and visualization effects.