Automatic Generation of 3-D Roof Training Dataset for Building Roof Segmentation From ALS Point Clouds

In the essential task of roof segmentation, the performance and generalization of deep learning-based methods for roof segmentation are much affected by the training datasets. Existing datasets for roof segmentation reveal the limitation of size and diversity. To address these issues, a framework is proposed to achieve fully automatic generation of point cloud roof segmentation datasets in this study. This framework fully leverages open 3-D building model data in LoD2, generating roof point clouds that contain real geographic information as well as accurate roof segment information. The point density and noise level can be defined by users in the framework, ensuring its flexibility and practicality across different environments. A pipeline is also proposed to utilize the generated point clouds for assisting in the training of deep learning-based roof segmentation methods. For validation purposes, a generated roof segmentation dataset including 50463 3-D building models, NRW3D, is created based on open geodata in North Rhine-Westphalia (NRW) state, Germany. The experimental results on NRW3D and the RoofNTNU dataset demonstrate the effectiveness and quality of the generated dataset as well as the proposed framework and pipeline. The source code for the proposed framework is available at Zenodo.