Large-scale 3D surface reconstruction at the city scale from UAV imagery is a core task for building digital-twin cities. Although 3D Gaussian Splatting (3DGS) and its variants have achieved remarkable progress in radiance-field reconstruction, existing methods still struggle with scale-varying structural complexity and diverse urban textures: geometry is often incomplete in low-frequency, weak-texture regions over large areas, while in high-frequency regions with intricate details, disordered primitive distributions tend to cause floating artifacts and outliers. To address these issues, we propose UrbanQGS, a large-scale urban reconstruction framework built on quadratic Gaussian splatting. We inherit quadratic surface Gaussians as scene primitives and use their analytic curvature as the central driving signal to perform a two-stage geometric refinement: (1) A self-supervised geometric constraint based on cross-view curvature consistency, which effectively fills geometric holes in weak-texture planar regions. (2) A curvature-aware adaptive representation allocation strategy that reweights, recalibrates, and ranks densification gradients, guiding split and clone operations toward geometrically complex regions while standard pruning removes redundant primitives. In addition, we integrate a monocular normal prior to correct global orientation drift during large-scale reconstruction. We further assess the compatibility of UrbanQGS with partitioned large-scene reconstruction using a visibility-aware block-wise protocol with hardware-constrained per-block budgets. The blocks are trained sequentially on a single GPU. Experiments on large-scale UAV datasets demonstrate that, compared with state-of-the-art 3DGS-based large-scale reconstruction methods, UrbanQGS maintains superior geometric quality while substantially reducing redundant Gaussians and mesh complexity, yielding finer and more complete geometric structures, while qualitative results on our custom-collected dataset indicate its compatibility with a partitioned large-scene reconstruction pipeline.
The automatic reconstruction of Building Information Models (BIMs) remains challenging, especially when applied to multi-storey buildings and indoor environments with non-Manhattan layouts and curved structures. To address these challenges, this paper presents a novel method that places Constructive Solid Geometry (CSG) at the core of building reconstruction from point clouds. The methodology also integrates the Segment Anything Model (SAM) and knowledge-based rules to reconstruct as-built BIMs while preserving the geometry and topology of building elements. The proposed approach consists of two main stages: SAM-based segmentation and CSG-based modelling. First, SAM is used to segment rooms, as well as door and window openings. Solids are then generated for the indoor spaces and openings. Finally, a set of Boolean operations is applied to reconstruct building elements, including slabs, walls, indoor spaces, doors, and windows, while subtracting openings of doors, windows, and staircases. Finally, the reconstructed model is encoded with semantic information following the Industry Foundation Classes (IFC) schema. The method was evaluated on seven two-storey building point clouds and demonstrated strong performance, achieving a mean Intersection over Union (mIoU) greater than 90% for floor plan reconstruction, a precision above 87% for door and window detection, and mean geometric deviations below 8 cm for BIM reconstruction.
