A way to reconstruct fine-grained building surfaces across a whole city with 3D Gaussian Splatting without downsampling the input imagery. Existing large-scale systems cut the scene into spatial blocks and shrink the images to fit GPU memory, which destroys the pixel-level guidance that building edges need, and they optimize building Gaussians together with vegetation, ground and sky, so foreground and background contaminate each other and edges come out wrong. This approach partitions semantically instead: a text-prompted segmentation foundation model produces coarse building masks on every registered image, a reliability-scored multi-view voting filter keeps only mask points confirmed across several views, and the camera visibility graph together with the camera-to-sparse-point correspondence groups the building area into sub-groups that each carry their own cameras, sparse points and masks and can be optimized independently and in parallel at full resolution. Two losses support it: a boundary-aware normal loss that down-weights the reference normal exactly where a mask boundary makes it meaningless, and a balance-load loss weighted by gradient magnitude so that thread load in tile-parallel rasterization is equalized against scene complexity rather than uniformly. Per-group meshes are fused with a truncated signed distance function and merged. The result beats 2DGS, SuGaR, GOF and PGSR on precision, recall and F1 across two GauU-Scene benchmarks and a self-collected dataset, with the load balancing alone cutting training time by roughly twelve percent.
Wang and colleagues at Wuhan University attack a specific failure of 3D Gaussian Splatting (3DGS) at city scale: reconstructing building surfaces in fine detail across a large urban area. Two obstacl…