GPUMLib: Deep Learning SOM Library for Surface Reconstruction

摘要

The evolution of 3D scanning devices and innovation in computerprocessing power and storage capacity has sparked the revolution ofproducing big point-cloud datasets. This phenomenon has becomingan integral part of the sophisticated building design processespecially in the era of 4th Industrial Revolution. The big point-clouddatasets have caused complexity in handling surface reconstructionand visualization since existing algorithms are not so readilyavailable. In this context, the surface reconstruction intelligentalgorithms need to be revolutionized to deal with big point-clouddatasets in tandem with the advancement of hardware processingpower and storage capacity. In this study, we propose GPUMLib –deep learning library for self-organizing map (SOM-DLLib) to solveproblems involving big point-cloud datasets from 3D scanningdevices. The SOM-DLLib consists of multiple layers for reducingand optimizing those big point cloud datasets. The findings show thefinal objects are successfully reconstructed with optimizedneighborhood representation and the performance becomes better asthe size of point clouds increases.