A BIM-Guided Virtual-to-Real Framework for Component-Level Semantic Segmentation of Construction Site Point Clouds.

Area of Science:

• Computer Vision and Machine Learning
• Geospatial Information Systems
• Construction Informatics

Background:

• LiDAR point cloud semantic segmentation is crucial for scan-to-BIM (Building Information Modeling) workflows.
• Current deep learning methods require extensive annotated real-world data, which is costly and difficult to obtain due to occlusions and noise.
• Existing approaches struggle with domain shift between synthetic and real-world data.

Purpose of the Study:

• To propose a novel BIM-guided Virtual-to-Real (V2R) framework for semantic segmentation of LiDAR point clouds.
• To eliminate the need for real-world annotations by training solely on synthetic data.
• To improve the accuracy and robustness of scan-to-BIM workflows in construction environments.

Main Methods:

• Generation of a large-scale Synthetic Point Cloud (SPC) dataset directly from BIM models with component-level labels.
• Development of a multi-feature fusion network combining Point Context Transformer (PCT) for global context and PointNet++ for local geometry.
• Implementation of a learnable point cloud augmentation module and multi-level domain adaptation strategies to address domain differences.

Main Results:

• The V2R framework achieved high performance on real construction data, with 70.89% overall accuracy and 53.14% mean IoU.
• The proposed Fusion model consistently outperformed baseline models on both scene-metric and component-level scores.
• Demonstrated stable performance across diverse real construction scenes and floors, validating the effectiveness of BIM-generated SPC.

Conclusions:

• The BIM-guided V2R framework successfully enables accurate semantic segmentation without real-world annotations.
• BIM-generated synthetic point clouds are effective for training robust deep learning models for construction applications.
• The framework shows significant potential for automated BIM-reality updates and site monitoring in construction.