A platform for research: civil engineering, architecture and urbanism
A platform for research: civil engineering, architecture and urbanism
Dual-PointNet
A two-stage deep learning network for semantic segmentation of point clouds
Bridges represent an essential element of the transportation network, but they face deterioration due to environmental and human factors. Building Information Modeling (BIM) can facilitate the maintenance planning of the bridges, but this requires digital building models of the existing structures, which are rarely available. These as-is models must be created retrospectively, e.g., from point clouds, which, if done manually, is a costly and time-consuming process. Thus, to partially automate this using deep learning is beneficial. In this paper, we propose a two-stage deep learning network that addresses the semantic segmentation of point clouds of bridges into individual components. Therefore, two instances of PointNet are combined, where the output from the first instance enriches the input of the second instance. The first PointNet is trained to segment the input point cloud into general classes, i.e., background, substructure, and superstructure. The second instance uses the enriched input point cloud to predict the individual bridge component classes, such as abutment, railing, or bridge cap. The combined network is trained with the accumulated loss of both sets of predictions. The proposed approach achieves promising results with a mIoU of 69% on point clouds from real-world railway bridges.
Dual-PointNet
A two-stage deep learning network for semantic segmentation of point clouds
Bridges represent an essential element of the transportation network, but they face deterioration due to environmental and human factors. Building Information Modeling (BIM) can facilitate the maintenance planning of the bridges, but this requires digital building models of the existing structures, which are rarely available. These as-is models must be created retrospectively, e.g., from point clouds, which, if done manually, is a costly and time-consuming process. Thus, to partially automate this using deep learning is beneficial. In this paper, we propose a two-stage deep learning network that addresses the semantic segmentation of point clouds of bridges into individual components. Therefore, two instances of PointNet are combined, where the output from the first instance enriches the input of the second instance. The first PointNet is trained to segment the input point cloud into general classes, i.e., background, substructure, and superstructure. The second instance uses the enriched input point cloud to predict the individual bridge component classes, such as abutment, railing, or bridge cap. The combined network is trained with the accumulated loss of both sets of predictions. The proposed approach achieves promising results with a mIoU of 69% on point clouds from real-world railway bridges.
Dual-PointNet
A two-stage deep learning network for semantic segmentation of point clouds
Mahameed, Ibrahim (author) / Faltin, Benedikt (author)
2023
277 KB
Miscellaneous
Electronic Resource
English
| DOAJ | 2023
Ausbildung - Dual ausbilden, dual studieren
| Online Contents | 2013