Urban-GS: A Unified 3D Gaussian Splatting Framework for Compact and High-Fidelity Aerial-to-Street Reconstruction

Paper | Demo | Checkpoints

CVPR 2026

Meng Wang^1,2   Changqun Xia²   Yuze Wang¹   Junyi Wang³   Wantong Duan^1,4   Xinxiong Xie²   Yue Qi^1,4

¹State Key Laboratory of Virtual Reality Technology and Systems, Beihang University
²PengCheng Laboratory   ³School of Computer Science and Technology, Shandong University
⁴Qingdao Research Institute of Beihang University

▶ Click to watch the demo video

Overview

This repository contains the implementation of Urban-GS, a unified 3D Gaussian Splatting framework for compact and high-fidelity aerial-to-street reconstruction. Urban-GS seamlessly models complex urban scenes, robustly supporting drastic viewpoint changes while maintaining visual fidelity across multi-scale observations. Compared to the state-of-the-art Horizon-GS, our approach delivers significant advantages in both average novel-view synthesis quality and representation compactness.

🛠️ Installation

# Clone the repository
git clone https://github.com/wangm-buaa/Urban-GS.git
cd Urban-GS

# Create conda environment
conda create -n urban-gs python=3.8 -y
conda activate urban-gs

# Install dependencies

# Replace the torch vision according your own hardware
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu118

pip install -r requirements.txt

pip install submodules/gsplat-urbangs
pip install submodules/fused-ssim
pip install submodules/simple-knn

🚀 Quick Start

Data Preparation

For benchmark evaluation, please follow the dataset preparation pipeline provided by Horizon-GS.

For custom aerial-to-street captures, Horizon-GS notes that scene initialization can be performed with the commercial software ContextCapture. Here, we additionally introduce several open-source SfM components. A practical setup is to use Hierarchical-Localization for feature extraction and matching, together with either Pixel-Perfect-SfM or GLOMAP for mapping. A lightweight alignment example is provided in preprocess/aerial_street_align.py.

The COLMAP 4.x series also supports modern pipelines such as ALIKED + LightGlue + GLOMAP. These components can be used as alternatives depending on the capture condition and reconstruction quality.

For depth and mask generation, please refer to the preprocessing tools released by hierarchical-3d-gaussians.

Training

Set source_path in the config file to the directory where your dataset is stored.

# Train on a scene with default settings
python train.py --config <path_to_config>

# Example: Train on Colosseum scene
python train.py --config config/ours/urbangs/synthetic/colosseum/urbangs.yaml

If memory is insufficient to load all images at once, enable --not_pre_load to load data dynamically during training.

Rendering

# Render novel views
python render.py -m <output_path> --skip_train

Evaluation

# Evaluate rendering quality
python metrics.py -m <output_path>

🖥️ Visualization

Our visualization tool is built on top of nerfstudio, with Urban-GS-specific modifications. For detailed usage instructions, please refer to the nerfstudio documentation and the instructions in the visualization tool repository.

Visualization_example.mp4

Note: The web viewer preview may exhibit lower visual quality than the actual rendered output video.

Acknowledgement

We gratefully acknowledge the following projects that have inspired and contributed to our work:

• Horizon-GS: Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes
• MaskGaussian: Adaptive 3D Gaussian Representation from Probabilistic Masks
• Octree-GS: Towards Consistent Real-time Rendering with LOD-Structured 3D Gaussians
• A Hierarchical 3D Gaussian Representation for Real-Time Rendering of Very Large Datasets
• Drone-assisted Road Gaussian Splatting with Cross-view Uncertainty
• gsplat
• nerfstudio

Citation

If you find this work useful in your research, please consider citing:

@InProceedings{Wang_2026_CVPR_Urban_GS,
    author    = {Wang, Meng and Xia, Changqun and Wang, Yuze and Wang, Junyi and Duan, Wantong and Xie, Xinxiong and Qi, Yue},
    title     = {Urban-GS: A Unified 3D Gaussian Splatting Framework for Compact and High-Fidelity Aerial-to-Street Reconstruction},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {33207-33216}
}

License

Please follow the LICENSE of 3D-GS.

Contact

• Meng Wang: wangm05@buaa.edu.cn