Deep Learning based Classification of Color Point Cloud for 3D Reconstruction of Interior Elements of Buildings

In architecture and engineering, the production of 3D models of various objects that are both simple and most closely related to reality is of particular importance. In this article, we are going to model different aspects of the interior of a building, which is performed in three general steps. In the first step, the existing point clouds of a room are semantically segmented using the PointNet Deep Learning Network. Each class of objects is then reconstructed using three methods including: Poisson, ball-pivoting and combined volumetric triangulation method and marching cubes. In the last step, each model is simplified by the methods of vertex clustering and edge collapse with quadratic error. Results are quantitatively and qualitatively evaluated for two types of objects, one with simple geometry and one with complex geometry. After selecting the optimal surface reconstruction method and simplifying it, all the objects are modeled. According to the results, the Poisson surface reconstruction method with a simplified edge collapse method provides better geometric accuracy of 0.1 mm for simpler geometry classes. In addition, for more complex geometry problems, the model produced by combined volumetric triangulation method and marching cubes with simplified edge collapse method was more suitable due to a higher accuracy of 0.022 mm.