Generation and Evaluation of Synthetic Point Cloud Data for Training of Machine Learning Models

Aug 3, 2026

Upload by Skyler

Machine learning is increasingly used to automate the creation of Building Information Models (BIM) and to perform point cloud semantic segmentation, yet training depends on large volumes of costly, time-consuming data capture and labeling. The paper presents a virtual laser scanning approach to generate synthetic point clouds as a cost-effective alternative to manual collection. Experiments train semantic segmentation models on a hybrid dataset combining real-world and synthetic scans, showing that hybrid-trained models outperform those trained only on real data.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages9

Reading Time23 min

AccessFree to download

Introduction

Motivation and digital representation

Scan-to-BIM and point cloud generation

Challenges in data capture and preprocessing

Proposed approach and contributions

Generation and Evaluation

Generation and Evaluation of Synthetic Point Cloud Data for Training of Machine Learning Models Hanxin Ma1, Jan Martens1, Jörg Blankenbach 1 1Chair of Computing in Civil Engineering and Geoinformation Systems and Geodetic Institute, RWTH Aachen University, Aachen, Germany ABSTRACT: Machine learning is increasingly applied for automating the creation of Building Information Models (BIM) of infrastructure objects and for point cloud semantic segmentation. However, machine learning is often costly and time-consuming due to the substantial volume of point cloud training data required which in turn must be captured and processed. This paper proposes the use of virtual laser scanning software to generate synthetic point clouds as a cost-effective alternative to manual capturing. It evaluates the impact of training machine learning models for infrastructure semantic segmentation using a hybrid dataset of real-world and synthetic point clouds. The findings indicate that models trained on hybrid datasets markedly outperform those trained solely on real-world data. This improvement illustrates the potential of augmenting datasets with synthetic data to not only improve model accuracy but also reduce the resources needed for manual data collection and labelling. KEYWORDS: Infrastructure, Synthetic point clouds, Scan-to-BIM, Machine Learning, Semantic Segmentation 1. INTRODUCTION Infrastructure objects such as bridges and roads are critically prone to deterioration over time, resulting in safety concerns that necessitate regular maintenance or reconstruction. In response, the architecture, engineering, and construction (AEC) industry increasingly relies on Building Information Modelling (BIM) to enhance the resilience of infrastructure objects (Eastman et al. , 2011) . In the realm of infrastructure management, the precise digital representation of infrastructure objects is fundamental to the creation of “as-is"BIM models. Advancing beyond these static models, the concept of digital twins marks a great evolution by providing a dynamic, real-time representation of these assets (Borrmann et al. , 2022),(Crampen and Blankenbach, 2023) . Scan-to-BIM technology is widely used to create accurate “as-is" BIM models of infrastructure objects, which are also essential for developing digital twins (Deng et al. , 2021) . This process integrates scanning technologies into the BIM framework, enabling the precise conversion of realworld data into digital BIM-enabled models (Son et al. , 2015) . Following the Scan-to-BIM paradigm , point clouds represents the first step in reconstruction. Point clouds are primarily obtained through two methods: 3D reconstruction from image data, utilizing techniques such as dense image matching (DIM) (Remondino et al. , 2013) and structure from motion (SfM) (Schonberger and Frahm, 2016) , and 3D laser scanning, which includes terrestrial (TLS) and mobile laser scanning (MLS) (Wang and Kim, 2019) . Although effective, these methods are often costly and time-intensive, particularly for extensive infrastructure data collection (Laefer, 2013),(Williams et al. , 2013), (Lu and Brilakis, 2017) . Additionally, the point clouds generated require substantial pre-processing, including registration (Yan et al. , 2017) and denoising (Zhou et al. , 2022) , to be usable for BIM models creation. The next step is semantic segmentation, where objects are segmented and classified based on geometric attributes. Manual segmentation is not only timeconsuming but also susceptible to errors. To overcome these limitations, current technology employs supervised machine learning techniques to automate this step (Perez-Perez et al. , 2021) . Machine learning models, trained on accurately labelled data, automate the conversion of 3D scan data into BIM elements, enhancing both the speed and accuracy of the process through continuous learning and algorithm optimization. These models depend on extensive training datase

FAQs

Why is synthetic point cloud data needed for training machine learning models?

Training requires substantial real-world point cloud data that is expensive and time-consuming to capture and label. Synthetic point clouds provide a more cost-effective data source while supporting dataset augmentation.

How are synthetic point clouds generated in the proposed approach?

Virtual laser scanning software is used to generate synthetic point clouds of infrastructure objects, avoiding labor-intensive manual capturing and related preprocessing steps.

What is the key finding from evaluating models trained on hybrid datasets?

Models trained on hybrid datasets that combine real-world and synthetic point clouds significantly outperform models trained solely on real-world data, indicating that synthetic data improves accuracy and reduces manual data requirements.

Aug 3, 2026

Upload by Skyler