Data & Code

Semantic 3D city models serve as a crucial geospatial framework, foundation, and vital spatial data source that underpins various smart city applications. While oblique photogrammetry has made significant progress in 3D construction, the models it generates continue to encounter issues such as high data volume, challenges in monomerization, and numerous geometric and texture defects. Particularly the use of intricate algorithms often leads to inefficiencies and high expenditures in city-scale modeling when introducing semantic information. Hence, this study proposes the SemCity-Gaussian method based on 3D Gaussian Splatting. It aims to perform high-precision 3D scene reconstruction and fast, high-quality rendering of typical urban geographic features with great significance while preserving the geographic semantics. Experimental results demonstrated that the models generated by the SemCity-Gaussian method exhibit outstanding performance in semantic segmentation and geometric accuracy, with an overall mean Intersection over Union (mIoU) exceeding 80% and a mean geometric error of only 0.008 meters compared to ground truth model obtained by 3D laser scanning. The SemCity-Gaussian method can generate semantic 3D models rapidly, accurately, and with low expenditure, providing a more efficient and intelligent solution for building Internet + 3D GIS platforms and promoting the construction of smart cities.

语义三维城市模型（Semantic 3D city models）是支撑各类智慧城市应用的核心地理空间框架、基础底座与关键空间数据源。尽管倾斜摄影测量（oblique photogrammetry）在三维建模领域已取得长足进展，但其生成的模型仍存在数据体量庞大、单体化难度高、几何与纹理缺陷较多等问题。尤其是在引入语义信息时，复杂算法的应用往往会导致城市级建模效率低下且成本高昂。为此，本研究提出了基于三维高斯溅射（3D Gaussian Splatting）的SemCity-Gaussian方法。该方法旨在在保留地理语义信息的前提下，实现高精度三维场景重建，并对典型城市地理要素实现快速高质量渲染，具备重要应用价值。实验结果表明，SemCity-Gaussian方法生成的模型在语义分割与几何精度方面表现优异：其总体平均交并比（mean Intersection over Union，mIoU）超过80%，与三维激光扫描（3D laser scanning）获取的真值模型相比，平均几何误差仅为0.008米。SemCity-Gaussian方法能够快速、精准且低成本地生成语义三维城市模型，为搭建互联网+三维地理信息系统（3D GIS）平台提供了更为高效智能的解决方案，同时助力智慧城市建设。