 This study develops a method called MS3DB for extracting multi-scale 3D building information from point cloud data, using surface flatness, variance in normal direction, and gray level co-occurrence matrix as labeling features. The method achieves high accuracy in extracting building labels at the object, grid, and block scales, with limited accuracy in extracting building edges but high accuracy in other parameters such as area, volume, and planar area index. The results show that integrating 2D and 3D features improves the accuracy of characterizing urban building structure at the block scale. This article was authored by Shesong Cao, Kihao Wang, Mingyi Du, and others.