Regularized three-dimensional Gaussians for large-scale building reconstruction with spatial prior

Existing large-scale building reconstruction methods based on Neural Radiance Field (NeRF) usually suffer from severe radiative blur and extremely long training time. The emergence of three-dimensional Gaussian Splatting (3DGS) can produce realistic rendering quality, while possessing much faster rendering speed than NeRF-based methods. However, for large-scale building reconstruction, 3DGS tends to produce blurred details and floaters due to the incomplete coverage of structural details by multi-view images and influence of illumination changes. This paper proposes a novel approach for high-quality 3D reconstruction and real-time rendering of large-scale buildings based on regularized 3D Gaussians and the building spatial prior. Considering the correlation of adjacent views, two regularizers are introduced for the predicted geometry and color to reduce artifacts caused by unseen viewpoints. Moreover, the spatial prior information of building boundaries is incorporated into the optimization process to eliminate floaters on the surface through the point cloud filtering, thereby improving the visual quality of the reconstructed model. Experiments on real-world and public datasets demonstrate that the proposed method outperforms existing methods based on NeRF and 3DGS for large-scale building reconstruction in terms of rendering quality and geometric accuracy.