Global and Uniform Point Cloud Completion With Density-Sensitive Transformer for Small-Scale 3-D Object Reconstruction

Incompleteness and irregularity are inherent challenges in 3-D point clouds, which hinder point-cloud-based 3-D object or scene understanding, especially for complicated industrial scenarios. Previous point cloud completion methods typically suffer from the nonuniform density of the restored point cloud. In this article, we proposed a global and uniform point cloud completion algorithm via a density-sensitive transformer for reconstructing complete and fine-grained 3-D small-scale industrial components. We first designed a position and feature encoding module to aggregate discriminative point cloud features. Then, we constructed a density-sensitive transformer structure with a coarse point cloud generator that allowed for recovering the global structure and uniform local details of the 3-D object. The experimental results on a real-world industrial component dataset and a public multiobject dataset demonstrated that our method achieved state-of-the-art performance with the former CD-l1 of 8.17 x 10(-3), avg-NUC of 0.61 and the latter CD-l1 of 5.33 x 10(-3), avg-NUC of 0.74. Our method better balanced the global point cloud integrity and density uniformity in a coarse-to-fine pipeline, which benefited high-quality 3-D object reconstruction. The method has been applied to practical aero-engine component reconstruction in the real-world scenario.