GS4Buildings: Building-Centric Reconstruction

• GS4Buildings is a collection of building-centered frameworks that integrate remote-sensing, Gaussian Splatting, and digital-twin techniques using explicit geometric priors.
• The prior-guided Gaussian Splatting method leverages LoD2 semantic building models to initialize and regularize 2D Gaussian Splatting for improved 3D surface reconstruction under occlusions.
• GS4Buildings also encompasses junction-based geometric extraction and single-building mesh pipelines, enhancing efficiency and accuracy in urban modeling applications.

GS4Buildings is a label applied to multiple building-centered computational frameworks spanning remote-sensing extraction, Gaussian Splatting-based 3D reconstruction, and digital-twin integration. Most explicitly, it denotes a prior-guided Gaussian Splatting method that uses low-level Level-of-Detail-2 semantic 3D building models to initialize and regularize 2D Gaussian Splatting for robust building surface reconstruction. Elsewhere, the same label is used for a geometry-driven building extractor based on junction saliency in very-high-spatial-resolution imagery, for a single-building mesh-extraction pipeline built from Google Earth Studio, SAM2, GroundingDINO, 2D Gaussian Splatting, and TSDF fusion, and for a digital-twin system that couples Gaussian Splatting with Google Maps Platform and multi-agent LLM modules (Zhang et al., 10 Aug 2025).

1. Scope, nomenclature, and conceptual range

A recurrent source of confusion is that GS4Buildings is not used in one uniquely fixed sense. The name appears across distinct technical settings, each centered on buildings but targeting different outputs, priors, and operating regimes.

Usage	Core formulation	arXiv id
Prior-guided reconstruction	LoD2-guided 2DGS for 3D building surface reconstruction	(Zhang et al., 10 Aug 2025)
Geometric building extraction	Junction-based geometric saliency and Geometric Building Index	(Xia et al., 2018)
Single-building mesh pipeline	Google Earth Studio, SAM2 + GroundingDINO, 2DGS+, TSDF fusion	(Gao et al., 2024)
Digital twin framework	Gaussian Splatting, Google Maps APIs, and multi-agent LLM analysis	(Gao et al., 9 Feb 2025)

The prior-guided reconstruction usage is the one for which GS4Buildings is the paper title itself. In that formulation, the method is motivated by the observation that vanilla 2D Gaussian Splatting and image-only pipelines leave holes in façades and roofs under frequent occlusion, whereas many cities already provide LoD2 semantic building models in B-Rep form. The method therefore uses semantic building models as strong geometric priors for both Gaussian initialization and optimization (Zhang et al., 10 Aug 2025).

A plausible implication is that GS4Buildings functions less as a single canonical algorithmic object than as a building-focused research designation for pipelines in which explicit structure, geometry, or geospatial context is injected into reconstruction or extraction.

2. Prior-guided Gaussian Splatting for 3D building reconstruction

In "GS4Buildings: Prior-Guided Gaussian Splatting for 3D Building Reconstruction" (Zhang et al., 10 Aug 2025), the central design choice is to remove dependence on a conventional SfM initialization and instead seed Gaussians directly from a low-level LoD2 semantic 3D building model. The method begins by uniformly sampling 3D points $\{p_i\}$ on the LoD2 mesh $\mathcal{M}$ via face-area-weighted random sampling. For each sampled point and camera, it computes an expected depth $d^{\text{exp}}_{i,j}=\|p_i-c_j\|_2$ and an actual intersection depth $d^{\text{int}}_{i,j}$ from ray-mesh casting. Visibility is then defined by

$$v_{i,j}=1 \quad \text{iff} \quad \left|d^{\text{int}}_{i,j}-d^{\text{exp}}_{i,j}\right|<\varepsilon,\qquad \varepsilon=0.05\mathrm{m}.$$

A point is retained only if it is visible in at least $k$ views, with $\sum_j v_{i,j}\ge k$ and typically $k=2$. The retained points are converted into the 2DGS input format and seed planar Gaussian splats over the building surfaces.

This SfM-free strategy is paired with dense geometric priors rendered from the aligned LoD2 model. For each calibrated camera, the method raycasts the mesh to generate per-view depth, normal, and validity-mask tensors:

$$(\mathbf{D}_j,\mathbf{N}_j,\mathbf{M}_j)=\mathcal{R}(\mathcal{M},K_j,T_j).$$

These priors give the Gaussian optimizer explicit per-pixel targets in building regions rather than relying only on photometric agreement. The paper describes this as depth and normal “prior” supervision and positions it as a mechanism for enforcing surface consistency and structural accuracy (Zhang et al., 10 Aug 2025).

The formulation is explicitly tied to 2D Gaussian Splatting. The paper notes that 2DGS is particularly suitable for surface reconstruction because of its flattened Gaussian representation and integrated normal regularization. GS4Buildings extends that base formulation without introducing an additional network architecture; it is described as a purely optimization-based enhancement built on the open-source 2DGS repository (Zhang et al., 10 Aug 2025).

3. Loss formulation, scheduling, and operating modes

The optimization objective augments the vanilla 2DGS losses with building-prior terms. The base terms are a photometric/color loss,

$$\mathcal{L}_c=\frac{1}{|\Omega|}\sum_{\mathbf{x}\in\Omega}\|\hat c(\mathbf{x})-c_{\text{gt}}(\mathbf{x})\|_2^2,$$

a depth-distortion regularizer $\mathcal{M}$0, and a normal-consistency regularizer,

$\mathcal{M}$1

GS4Buildings adds a depth prior

$\mathcal{M}$2

and a normal prior

$\mathcal{M}$3

The total objective is

$\mathcal{M}$4

Here, $\mathcal{M}$5 and $\mathcal{M}$6 are rendered from the current Gaussian model, and $\mathcal{M}$7 is a learned scale factor (Zhang et al., 10 Aug 2025).

The schedule follows two phases. During iterations $\mathcal{M}$8–$\mathcal{M}$9, $d^{\text{exp}}_{i,j}=\|p_i-c_j\|_2$0 and $d^{\text{exp}}_{i,j}=\|p_i-c_j\|_2$1 are strongly weighted to enforce global completeness. During iterations $d^{\text{exp}}_{i,j}=\|p_i-c_j\|_2$2–$d^{\text{exp}}_{i,j}=\|p_i-c_j\|_2$3, the prior weights are gradually decayed and $d^{\text{exp}}_{i,j}=\|p_i-c_j\|_2$4 are ramped up to refine local smoothness and fine detail. An example schedule is $d^{\text{exp}}_{i,j}=\|p_i-c_j\|_2$5 always, $d^{\text{exp}}_{i,j}=\|p_i-c_j\|_2$6, $d^{\text{exp}}_{i,j}=\|p_i-c_j\|_2$7, and $d^{\text{exp}}_{i,j}=\|p_i-c_j\|_2$8, $d^{\text{exp}}_{i,j}=\|p_i-c_j\|_2$9 (Zhang et al., 10 Aug 2025).

The method also exposes two operating modes. In building-only mode, sampling and training are masked by $d^{\text{int}}_{i,j}$0 so that only building pixels contribute; in building-enhanced mode, the full scene is trained, but the prior losses still apply only in building regions. The building-only configuration is not merely a cropping heuristic: it changes the active primitive set and reduces the total number of Gaussian primitives by $d^{\text{int}}_{i,j}$1, which the paper links to improved efficiency and compactness (Zhang et al., 10 Aug 2025).

4. Quantitative performance, implementation details, and limitations

The main quantitative evaluation is conducted on the TUM2TWIN urban benchmark across $d^{\text{int}}_{i,j}$2 building scenes. For 3D accuracy, GS4Buildings reports a Chamfer Distance of $d^{\text{int}}_{i,j}$3 and an M3C2 score of $d^{\text{int}}_{i,j}$4, compared with $d^{\text{int}}_{i,j}$5 and $d^{\text{int}}_{i,j}$6 for vanilla 2DGS and $d^{\text{int}}_{i,j}$7 and $d^{\text{int}}_{i,j}$8 for MVS (Pix4Dmatic). The reported reductions are $d^{\text{int}}_{i,j}$9 in Chamfer versus 2DGS, $$v_{i,j}=1 \quad \text{iff} \quad \left|d^{\text{int}}_{i,j}-d^{\text{exp}}_{i,j}\right|<\varepsilon,\qquad \varepsilon=0.05\mathrm{m}.$$0 versus MVS, and $$v_{i,j}=1 \quad \text{iff} \quad \left|d^{\text{int}}_{i,j}-d^{\text{exp}}_{i,j}\right|<\varepsilon,\qquad \varepsilon=0.05\mathrm{m}.$$1 in M3C2 versus 2DGS (Zhang et al., 10 Aug 2025).

For 3D completeness, the threshold-based $$v_{i,j}=1 \quad \text{iff} \quad \left|d^{\text{int}}_{i,j}-d^{\text{exp}}_{i,j}\right|<\varepsilon,\qquad \varepsilon=0.05\mathrm{m}.$$2 score rises to $$v_{i,j}=1 \quad \text{iff} \quad \left|d^{\text{int}}_{i,j}-d^{\text{exp}}_{i,j}\right|<\varepsilon,\qquad \varepsilon=0.05\mathrm{m}.$$3 and VOC to $$v_{i,j}=1 \quad \text{iff} \quad \left|d^{\text{int}}_{i,j}-d^{\text{exp}}_{i,j}\right|<\varepsilon,\qquad \varepsilon=0.05\mathrm{m}.$$4, compared with $$v_{i,j}=1 \quad \text{iff} \quad \left|d^{\text{int}}_{i,j}-d^{\text{exp}}_{i,j}\right|<\varepsilon,\qquad \varepsilon=0.05\mathrm{m}.$$5 and $$v_{i,j}=1 \quad \text{iff} \quad \left|d^{\text{int}}_{i,j}-d^{\text{exp}}_{i,j}\right|<\varepsilon,\qquad \varepsilon=0.05\mathrm{m}.$$6 for 2DGS and $$v_{i,j}=1 \quad \text{iff} \quad \left|d^{\text{int}}_{i,j}-d^{\text{exp}}_{i,j}\right|<\varepsilon,\qquad \varepsilon=0.05\mathrm{m}.$$7 and $$v_{i,j}=1 \quad \text{iff} \quad \left|d^{\text{int}}_{i,j}-d^{\text{exp}}_{i,j}\right|<\varepsilon,\qquad \varepsilon=0.05\mathrm{m}.$$8 for MVS. The paper reports a $$v_{i,j}=1 \quad \text{iff} \quad \left|d^{\text{int}}_{i,j}-d^{\text{exp}}_{i,j}\right|<\varepsilon,\qquad \varepsilon=0.05\mathrm{m}.$$9 improvement in threshold-based completeness over 2DGS, $k$0 over MVS, and a $k$1 VOC improvement over 2DGS. In novel-view synthesis, the reported averages are PSNR $k$2, SSIM $k$3, and LPIPS $k$4 for GS4Buildings, compared with PSNR $k$5, SSIM $k$6, and LPIPS $k$7 for 2DGS (Zhang et al., 10 Aug 2025).

The implementation remains close to standard 2DGS toolchains. The paper specifies $k$8 total iterations, split evenly across the two phases; Trimesh for mesh sampling; Open3D for raycasting; visibility threshold $k$9; and $\sum_j v_{i,j}\ge k$0 as the visibility count. Gaussians are initialized from retained mesh samples and flattened into anisotropic discs in the local LoD2 tangent plane. Full-scene GS4Buildings uses approximately $\sum_j v_{i,j}\ge k$1 Gaussians, while building-only GS4Buildings uses $\sum_j v_{i,j}\ge k$2. Building-only mode runs approximately $\sum_j v_{i,j}\ge k$3 faster per iteration and uses approximately $\sum_j v_{i,j}\ge k$4 less GPU memory than full-scene GS4Buildings; both still train in approximately $\sum_j v_{i,j}\ge k$5 hours on a single NVIDIA A6000 (Zhang et al., 10 Aug 2025).

The limitations are stated directly. Fine decorative details such as windowsills and cornices can be oversmoothed by planar LoD2 priors. The method requires reasonably accurate LoD2 models in the target area and heavily relies on correct camera calibration, with no SfM fallback. The cited applications are smart-city analytics, digital twins, and urban planning and historic preservation (Zhang et al., 10 Aug 2025).

5. Other frameworks carrying the GS4Buildings label

GeoSay formulation. In the remote-sensing setting, GS4Buildings denotes GeoSay’s geometry-centered building extraction strategy for very-high-spatial-resolution imagery. The method extracts anisotropic-scale junctions, models an informative $\sum_j v_{i,j}\ge k$6-junction as $\sum_j v_{i,j}\ge k$7, computes first-order saliency

$\sum_j v_{i,j}\ge k$8

and second-order saliency by aggregating nearby junctions. The pixel-wise Geometric Building Index is then

$\sum_j v_{i,j}\ge k$9

Shadow responses are attenuated with a black top-hat transform and a small Gaussian blur may be applied. Reported results include mAP $k=2$0, $k=2$1 $k=2$2 on SpaceNet-65; mAP $k=2$3, $k=2$4 $k=2$5 on Massachusetts; and mAP $k=2$6, $k=2$7 $k=2$8 on Potsdam, with stronger cross-domain generalization than the cited HF-FCN baseline when that baseline is not trained on the target domain (Xia et al., 2018).

Single-building mesh extraction formulation. In "Gaussian Building Mesh (GBM): Extract a Building's 3D Mesh with Google Earth and Gaussian Splatting" (Gao et al., 2024), GS4Buildings names a five-stage pipeline: multi-view image acquisition via Google Earth Studio; text- or click-based masking with SAM2 + GroundingDINO; mask refinement through morphology and contour simplification; 2D Gaussian Splatting for radiance and depth estimation; and TSDF fusion followed by Marching Cubes. Typical settings are $k=2$9 frames, altitudes $$(\mathbf{D}_j,\mathbf{N}_j,\mathbf{M}_j)=\mathcal{R}(\mathcal{M},K_j,T_j).$$0–$$(\mathbf{D}_j,\mathbf{N}_j,\mathbf{M}_j)=\mathcal{R}(\mathcal{M},K_j,T_j).$$1, tilts $$(\mathbf{D}_j,\mathbf{N}_j,\mathbf{M}_j)=\mathcal{R}(\mathcal{M},K_j,T_j).$$2–$$(\mathbf{D}_j,\mathbf{N}_j,\mathbf{M}_j)=\mathcal{R}(\mathcal{M},K_j,T_j).$$3, and image resolution approximately $$(\mathbf{D}_j,\mathbf{N}_j,\mathbf{M}_j)=\mathcal{R}(\mathcal{M},K_j,T_j).$$4. The 2DGS+ training loss is

$$(\mathbf{D}_j,\mathbf{N}_j,\mathbf{M}_j)=\mathcal{R}(\mathcal{M},K_j,T_j).$$5

with $$(\mathbf{D}_j,\mathbf{N}_j,\mathbf{M}_j)=\mathcal{R}(\mathcal{M},K_j,T_j).$$6 and $$(\mathbf{D}_j,\mathbf{N}_j,\mathbf{M}_j)=\mathcal{R}(\mathcal{M},K_j,T_j).$$7. The pipeline typically densifies from approximately $$(\mathbf{D}_j,\mathbf{N}_j,\mathbf{M}_j)=\mathcal{R}(\mathcal{M},K_j,T_j).$$8 Gaussians to approximately $$(\mathbf{D}_j,\mathbf{N}_j,\mathbf{M}_j)=\mathcal{R}(\mathcal{M},K_j,T_j).$$9 by iteration $$\mathcal{L}_c=\frac{1}{|\Omega|}\sum_{\mathbf{x}\in\Omega}\|\hat c(\mathbf{x})-c_{\text{gt}}(\mathbf{x})\|_2^2,$$0, then extracts a mesh through Open3D TSDF fusion at roughly $$\mathcal{L}_c=\frac{1}{|\Omega|}\sum_{\mathbf{x}\in\Omega}\|\hat c(\mathbf{x})-c_{\text{gt}}(\mathbf{x})\|_2^2,$$1 voxel size. Against GS2Mesh on seven Google Earth Studio scenes, reported view-synthesis scores are PSNR $$\mathcal{L}_c=\frac{1}{|\Omega|}\sum_{\mathbf{x}\in\Omega}\|\hat c(\mathbf{x})-c_{\text{gt}}(\mathbf{x})\|_2^2,$$2 versus $$\mathcal{L}_c=\frac{1}{|\Omega|}\sum_{\mathbf{x}\in\Omega}\|\hat c(\mathbf{x})-c_{\text{gt}}(\mathbf{x})\|_2^2,$$3, SSIM $$\mathcal{L}_c=\frac{1}{|\Omega|}\sum_{\mathbf{x}\in\Omega}\|\hat c(\mathbf{x})-c_{\text{gt}}(\mathbf{x})\|_2^2,$$4 versus $$\mathcal{L}_c=\frac{1}{|\Omega|}\sum_{\mathbf{x}\in\Omega}\|\hat c(\mathbf{x})-c_{\text{gt}}(\mathbf{x})\|_2^2,$$5, and LPIPS $$\mathcal{L}_c=\frac{1}{|\Omega|}\sum_{\mathbf{x}\in\Omega}\|\hat c(\mathbf{x})-c_{\text{gt}}(\mathbf{x})\|_2^2,$$6 versus $$\mathcal{L}_c=\frac{1}{|\Omega|}\sum_{\mathbf{x}\in\Omega}\|\hat c(\mathbf{x})-c_{\text{gt}}(\mathbf{x})\|_2^2,$$7; average 3D-SSIM is $$\mathcal{L}_c=\frac{1}{|\Omega|}\sum_{\mathbf{x}\in\Omega}\|\hat c(\mathbf{x})-c_{\text{gt}}(\mathbf{x})\|_2^2,$$8 versus $$\mathcal{L}_c=\frac{1}{|\Omega|}\sum_{\mathbf{x}\in\Omega}\|\hat c(\mathbf{x})-c_{\text{gt}}(\mathbf{x})\|_2^2,$$9 (Gao et al., 2024).

Digital-twin formulation. In "Digital Twin Buildings: 3D Modeling, GIS Integration, and Visual Descriptions Using Gaussian Splatting, ChatGPT/Deepseek, and Google Maps Platform" (Gao et al., 9 Feb 2025), GS4Buildings is an end-to-end single-building digital-twin system. It combines Gaussian Splatting-based mesh reconstruction, Google Maps Platform APIs, and four kinds of agents—GPT-4o-latest, GPT-4o-mini, Deepseek-chat (V3), and Deepseek-reasoner (R1)—for view-level keyword extraction, aggregation, and caption generation. The framework reports median perplexities of approximately $\mathcal{M}$00 for GPT-4o-latest and approximately $\mathcal{M}$01 for GPT-4o-mini on $\mathcal{M}$02 image-to-keyword calls, and average CLIP scores of approximately $\mathcal{M}$03, $\mathcal{M}$04, $\mathcal{M}$05, and $\mathcal{M}$06 for GPT-4o-mini, ChatGPT-4o, Deepseek-chat, and Deepseek-reasoner, respectively, over $\mathcal{M}$07 captions. It also reports round-trip API latency for Geocoding + Elevation + Static Maps of approximately $\mathcal{M}$08 per building and a Google Maps overlay running at $\mathcal{M}$09 in a modern browser (Gao et al., 9 Feb 2025).

These usages are technically related through their building-centric focus, but they are not interchangeable. They target different outputs: 2D footprint extraction, single-building watertight meshes, GIS-linked digital twins, or LoD2-prior-guided urban building reconstruction.

6. Position within adjacent building and urban reconstruction research

GS4Buildings sits within a broader lineage of building extraction and structured urban modeling. An important precursor in aerial-scene building extraction is "Automatic Building Extraction in Aerial Scenes Using Convolutional Networks" (Yuan, 2016), which uses a seven-stage ConvNet followed by a branch-and-merge stage for pixel-wise prediction and introduces the signed distance function

$\mathcal{M}$10

The method leverages GIS footprints to compile training data, corrects raster-vector misalignment via cross-correlation between image gradients and polygon masks, and reports Precision $\mathcal{M}$11, Recall $\mathcal{M}$12 on the Washington, D.C. test set, with cross-city Precision $\mathcal{M}$13 and Recall $\mathcal{M}$14 (Yuan, 2016). This establishes a clear precedent for using authoritative geospatial building data as supervision or alignment prior.

That precedent is reinforced by work on open building datasets. "Open government geospatial data on buildings for planning sustainable and resilient cities" identifies more than $\mathcal{M}$15 releases from $\mathcal{M}$16 countries containing above $\mathcal{M}$17 million buildings and benchmarks them across five dimensions: accessibility, richness, data quality, harmonisation, and relationships with other actors (Biljecki et al., 2021). A plausible implication is that such authoritative footprint repositories provide the institutional and data-infrastructure substrate on which some GS4Buildings variants can build, especially when LoD models, footprints, height attributes, or official identifiers are required.

At the urban-scene level, the closest neighboring formulation is "GS4City: Hierarchical Semantic Gaussian Splatting via City-Model Priors" (Zhang et al., 13 Apr 2026). GS4City uses aligned LoD3 CityGML models, two-pass raycasting, parent-child semantic validation, multi-source mask fusion, and Gaussian identity encoding to transfer hierarchical building semantics into a photorealistic Gaussian scene representation. On TUM2TWIN and Gold Coast, it reports gains of up to $\mathcal{M}$18 IoU points in coarse building segmentation and $\mathcal{M}$19 mIoU points in fine-grained semantic segmentation over LangSplat and Gaga (Zhang et al., 13 Apr 2026). This suggests that the building-model prior paradigm associated with GS4Buildings extends naturally from geometry completion to hierarchy-aware semantic urban reconstruction.

Taken together, these adjacent works clarify what GS4Buildings is and is not. It is not a single universally standardized algorithm; rather, it is a family of building-centered formulations in which explicit priors—junction geometry, GIS footprints, LoD2 or LoD3 city models, segmentation masks, or cloud geospatial services—are coupled with dense prediction or Gaussian Splatting to improve extraction, reconstruction, or semantic integration.