Linketic/CityGaussian

What it solves

CityGaussian addresses the challenge of reconstructing high-quality, large-scale 3D scenes (such as city-scale environments) using Gaussian Splatting. It overcomes limitations related to memory costs and GPU constraints when dealing with massive datasets, while improving geometric accuracy and rendering speed.

How it works

The project implements a series of methods (V1 and V2) that utilize Gaussian Splatting for scene reconstruction. Key mechanisms include:

• Multi-GPU Reconstruction: A system for distributing the reconstruction process across multiple GPUs with controllable memory costs.
• Model Partitioning: Analysis and assignment of data to specific model partitions to handle large-scale scenes.
• Joint Optimization: The ability to perform joint pose and 3DGS optimization, which is particularly useful for imperfect camera pose data (e.g., results from 3D foundation models like VGGT-X).
• Mesh Extraction: Integration of 2DGS-style mesh extraction to convert Gaussian representations into geometric meshes.

Who it’s for

This tool is designed for researchers and developers working in 3D computer vision, urban planning, and digital twin creation who need to actually reconstruct large-scale real-world environments from images.

Highlights

• Scalable Infrastructure: Supports multi-GPU reconstruction with no limit on the number of GPUs.
• Geometric Accuracy: Focuses on geometrically accurate reconstruction, as highlighted in CityGaussianV2.
• Flexible Optimization: Supports joint pose and 3DGS optimization for noisy COLMAP results.
• Integrated Tooling: Includes a web viewer, trajectory-aligned rendering, and mesh video generation.