3D point cloud segmentation is a key task in urban scene reconstruction, especially for extracting building structures, which are diverse in scale and geometry. Existing segmentation methods mainly rely on supervised deep learning, which suffers from limited generalization across different scenes and requires large amounts of annotated 3D data and computational resources. In contrast, 2D image segmentation has achieved significant progress. This work proposes a generalized 3D building segmentation framework based on 2D–3D fusion. By leveraging state-of-the-art 2D segmentation models such as Mask2Former and SAM, and combining them with 3D point clouds reconstructed by COLMAP, we establish correspondences between 2D masks and 3D points. This approach enables effective segmentation of 3D buildings without 3D supervision, and lays a foundation for downstream tasks such as urban scene reconstruction, measurement, and mapping.