Inpainting occlusion holes in 3D built environment point clouds

Abstract

Point clouds obtained from mobile and terrestrial laser scanning are imperfect as data is typically missing due to occlusions. This problem is often encountered in 3D reconstruction and is especially troublesome for 3D visualization applications. The missing data may be recovered by intensifying the scanning mission, which may be expensive, or to some extent, by computational means. Here, we present an inpainting technique that covers these occlusion holes in 3D built environment point clouds. The proposed technique uses two neural networks with an identical architecture, applied separately for geometry and colors.