A platform for research: civil engineering, architecture and urbanism
A platform for research: civil engineering, architecture and urbanism
Enhancing Point Cloud Segmentation of Chinese Historical Buildings with Synthetic Data
Considering the challenges of segmenting architectural components in point cloud data, particularly for Chinese historical buildings, we develop an efficient method for constructing datasets to enhance deep learning techniques for precise segmentation. Surface sampling is integrated with advanced virtual laser scanning technology in our approach. Initially, labeled point cloud data is captured through surface sampling. Subsequently, the HELIOS++ simulation platform mimics real-world scanning to generate unlabeled data resembling actual point clouds. Precise alignment and label transfer between these two types of data result in an annotated dataset that preserves authentic scanning characteristics. Additionally, we introduce a symmetry-axis-based point cloud completion technique to address data loss during scanning, leveraging the inherent symmetry found in Chinese historical buildings. To validate the effectiveness of our dataset, two state-of-the-art deep learning models are selected for comprehensive evaluation. Experimental results demonstrate that our dataset supports efficient and stable model training, exhibits strong generalization capabilities, and provides a robust foundation for semantic segmentation of historical buildings.
Enhancing Point Cloud Segmentation of Chinese Historical Buildings with Synthetic Data
Considering the challenges of segmenting architectural components in point cloud data, particularly for Chinese historical buildings, we develop an efficient method for constructing datasets to enhance deep learning techniques for precise segmentation. Surface sampling is integrated with advanced virtual laser scanning technology in our approach. Initially, labeled point cloud data is captured through surface sampling. Subsequently, the HELIOS++ simulation platform mimics real-world scanning to generate unlabeled data resembling actual point clouds. Precise alignment and label transfer between these two types of data result in an annotated dataset that preserves authentic scanning characteristics. Additionally, we introduce a symmetry-axis-based point cloud completion technique to address data loss during scanning, leveraging the inherent symmetry found in Chinese historical buildings. To validate the effectiveness of our dataset, two state-of-the-art deep learning models are selected for comprehensive evaluation. Experimental results demonstrate that our dataset supports efficient and stable model training, exhibits strong generalization capabilities, and provides a robust foundation for semantic segmentation of historical buildings.
Enhancing Point Cloud Segmentation of Chinese Historical Buildings with Synthetic Data
Advances in Engineering res
Qiu, Yanjun (editor) / Feng, Weimin (editor) / Zhang, Zhiqiang (editor) / Ahmad, Fauziah (editor) / Ma, Jiangping (author) / Guo, Zhiyuan (author) / Li, Wenpeng (author) / Chen, Weiya (author)
International Conference on Civil Architecture, Hydropower and Engineering Management ; 2024 ; Kunming, China
2025-01-31
8 pages
Article/Chapter (Book)
Electronic Resource
English
GRAPH CNN with RADIUS DISTANCE for SEMANTIC SEGMENTATION of HISTORICAL BUILDINGS TLS POINT CLOUDS
| BASE | 2020
Instance Segmentation of Industrial Point Cloud Data
| ASCE | 2021
Structural Behaviors of Chinese Historical Timber-Frame Buildings
| Trans Tech Publications | 2012