User: Guest  Login
Document type:
Masterarbeit
Author(s):
Kirn, Hannes
Title:
Reconstruction of truncated instance point clouds with the help of generative models
Abstract:
Currently, Point Clouds (PCs) are widely used especially in the Architecture, Engineering and Construction (AEC) industry for various use cases, such as planning equipment relocation in a factory; layout, and assembly line planning; global manufacturing operations; 5S and Gemba Walks; best practice sharing; visual interfaces, and in general Scan-to-BIM. Therefore, an accurate PC is very useful for getting the best information out of it and will be more helpful in processing the use cases more efficiently. This Master Thesis aims to improve PCs especially in Mechanical, Electrical and Plumbing (MEP) environments with the use of Machine Learning (ML) algorithms such as Voxel-Deep Convolutional Generative Adversarial Networks (Voxel-DCGAN) and Point Completion Networks (PCNs). For this procedure, synthetic data is firstly generated as training datasets before completing the real truncated PC clusters. An evaluation is performed afterwards for the applied methods. For the procedure without pre-voxelization, the assignments of different materials were modified for each instance to later be able to segment each PC instances in the output of the simulations. Afterwards, a PCN was applied to complete those components. For applying the Voxel-DCGAN method, the voxelization step on the PC is needed. This can be done by the given code from MIN (2004); (NOORUDDIN & TURK, 2003). The given Voxel-DCGAN method which contains a 3D shape generative model can be used for completion.
Subject:
ALL Allgemeines
Supervisor:
Collins, Fiona and Nousias, Stavros and Hu Zhiqi and Borrmann, Andre
Year:
2024
Year / month:
2024-04
Month:
Apr
University:
Technische Universität München
 BibTeX