A Computer Vision Approach for Building Facade Component Segmentation on 3D Point Cloud Models Reconstructed by Aerial Images

Segmenting windows and doors on 3D point cloud models allows for heat loss audits around these areas. Researchers have collected aerial images to reconstruct 3D models for large districts, but easily accessible training datasets with data acquired on ground level cannot be directly used for segmentation on 3D models reconstructed by aerial images. Additionally, building a new dataset is a time-consuming and labour-intensive process. Therefore, we propose a segmentation approach that uses open source training datasets to segment windows and doors on façade images rendered from 3D point clouds. The results show that our approach can make full use of open source
datasets to segment windows and doors, and that such trained segmentation models performs differently for different building styles. In addition, different algorithms result in various degrees of accuracy and segmentation on windows performs better than on doors.