Research map / 3D vision / Gaussian splatting

Horizon-GS: Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes

CVPR 2025 · 3D vision › Gaussian splatting · approach: Compact and efficient Gaussian splatting · 13 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Horizon-GS unifies aerial and street view reconstruction with Gaussian splatting via a coarse-to-fine two-stage training, balanced camera sampling, and a new partly synthetic, partly real aerial-street dataset.

Problem

Unified reconstruction and rendering of large-scale scenes from both aerial and street views is hard because conflicting densification gradients and imbalanced camera coverage degrade single-model Gaussian splatting, and calibrated cross-view datasets are scarce.

Approach

A two-stage coarse-to-fine training: stage one builds coarse geometry from aerial views only, stage two freezes the MLPs and refines with street views using a new significance-based densification criterion, combined with a view-sampling ratio that balances aerial and street images and a view-dependent multi-level LOD scheme built on Scaffold-GS anchors.

Key results

• On the synthetic Block Small scene, Horizon-GS reaches 30.59 PSNR for aerial views versus 28.44 for Scaffold-GS and 25.44 for 3D-GS.
• On the large-scale Hillside Summer scene, Horizon-GS reaches 31.09 aerial PSNR versus 25.46 for Scaffold-GS* and 24.89 for Vast-GS.
• Horizon-GS achieves real-time rendering at 51.5 FPS on the Hillside Morning and Hillside Summer scenes, with about four hours of training per scene.

Measured effects reported in the paper

Surprising findings

• Joint training of aerial and street views underperforms separate training, with fewer final Gaussian primitives as more views are added.
• Hierarchical-3DGS remains competitive on street views in several scenes, attributed to its street-focused design.

Limitations

• Relies on dense multi-view input and depth supervision from pretrained monocular depth or rendering engines.
• Real-data quality is lower than synthetic due to pose estimation, sensor noise, and texture complexity.
• Assumes Manhattan world with z-axis perpendicular to ground.

Experimental setup

Datasets: MatrixCity UC-GS dataset Horizon-GS captured aerial-street dataset

Benchmarks: Block Small Block A Hillside Morning Hillside Summer UC-GS NYC and SF Real captured Road and Park scenes

Metrics: PSNR SSIM LPIPS FPS

Baselines: 3D-GS Scaffold-GS Hierarchical-3DGS 2D-GS UC-GS Vast-GS City-GS

Builds on: 3D Gaussian Splatting Scaffold-GS 2D Gaussian Splatting Octree-GS VastGaussian

Closest papers by meaning

• MS-GS: Multi-Appearance Sparse-View 3D Gaussian Splatting in the Wild (NeurIPS 2025 · similarity 0.91)
• Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.90)
• EVolSplat: Efficient Volume-based Gaussian Splatting for Urban View Synthesis (CVPR 2025 · similarity 0.90)
• GS-IR: 3D Gaussian Splatting for Inverse Rendering (CVPR 2024 · similarity 0.90)
• SplatFormer: Point Transformer for Robust 3D Gaussian Splatting (ICLR 2025 · similarity 0.90)
• PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)