University of Twente Student Theses


Training deep networks with BIM models for indoor point cloud classification

Muhammad Nuzul, Mahfiruddin Syah Dema (2023) Training deep networks with BIM models for indoor point cloud classification.

[img] PDF
Abstract:Deep learning methods has been used in the point cloud classification applications. Particularly, it is used to provide as-built conditions of the buildings for construction progress monitoring. However, there is limited availability of labeled indoor point cloud datasets publicly available to train the deep learning network. Consequently, it can brings incorrect information and lead to cost overrun. Nevertheless, Building Information Models or BIM are available as it is used as the design model for the buildings. Therefore, this research leverages the BIM models to generate synthetic point clouds that can overcome this problem. The main results of this research is that this approach can successfully generate the synthetic point clouds to be used as additional dataset for point clouds classification. The networks trained on the synthetic point clouds has 14.22% mean – Intersection over Union (m-IoU) differences compared to the benchmark point clouds dataset, the S3DIS. Additionally, by augmenting the synthetic point clouds and the S3DIS dataset, it has 17.69% m-IoU differences compared to only using the S3DIS dataset. However, this approach failed completely classify stair and window elements due to class-imbalance and inter-class similarity problems.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Subject:38 earth sciences
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page