Research map / 3D vision / Gaussian splatting
CityGaussianV2: Efficient and Geometrically Accurate Reconstruction for Large-Scale Scenes
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.
Scales 2D Gaussian Splatting to large-scale scenes with decomposed-gradient densification, depth regression from Depth-Anything-V2, an elongation filter, and a streamlined CityGaussian-style parallel training and compression pipeline, plus a TnT-style geometry evaluation protocol.
Problem
Accurate surface geometry and efficient parallel training/compression for large-scale scene reconstruction with Gaussian splatting remain hard: 2DGS suffers slow convergence, blurry surfels, and Gaussian count explosion under parallel training, and existing geometry benchmarks for large scenes are u
Approach
Use 2DGS as primitive, densify using only the SSIM-loss gradient (decomposed-gradient densification) to remove blurry surfels, filter densification of extremely elongated surfels to prevent count explosion, supervise with monocular depth, and replace CityGaussian's post-pruning and distillation with contribution-based trimming during block tuning and SH degree 2 from scratch.
Key results
- On GauU-Scene, the method reaches F1 0.501 versus 0.453 for CityGS and 0.491 for 2DGS, with PSNR 24.51 versus 24.75 for CityGS.
- On MatrixCity-Aerial, F1 is 0.556 versus 0.462 for CityGS, with recall 0.752 versus 0.637.
- On Residence ablation, parallel tuning with trim cuts memory from 31.5 GB to 19.0 GB and storage to 1.90 GB, and the SH-degree-2 plus VQ small version reaches 0.44 MB... reported as 0.44 GB storage and 14.2 GB memory.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CityGaussianV2 instead of CityGS | 2DGS-based reconstruction · GauU-Scene | F1: 0.453 -> 0.501 | Helped |
| Use CityGaussianV2 instead of CityGS | 2DGS-based reconstruction · MatrixCity-Aerial | F1: 0.462 -> 0.556 | Helped |
| Use CityGaussianV2 instead of CityGS | 2DGS-based reconstruction · MatrixCity-Street | F1: 0.401 -> 0.503 | Helped |
| Add the Elongation Filter (ablation) | 2DGS pretraining · Residence (GauU-Scene) | F1: 0.410 -> 0.411 | No clear effect |
| Add DGD densification (ablation) | 2DGS with Elongation Filter · Residence (GauU-Scene) | PSNR: 21.18 -> 22.24 | Helped |
| Add Depth Regression (ablation) | 2DGS with Elongation Filter and DGD · Residence (GauU-Scene) | F1: 0.429 -> 0.438 | Helped |
| Add parallel tuning (ablation) | CityGaussianV2 · Residence (GauU-Scene) | F1: 0.438 -> 0.471 | Helped |
| Add contribution-based trimming (ablation) | parallel-tuned CityGaussianV2 · Residence (GauU-Scene) | F1: 0.471 -> 0.477 | Helped |
| Replace contribution-based trimming with pruning (ablation) | parallel-tuned CityGaussianV2 · Residence (GauU-Scene) | SSIM: 0.742 -> 0.738 | Hurt |
| Use SH degree 2 from the start (ablation) | CityGaussianV2 · Residence (GauU-Scene) | PSNR: 23.50 -> 23.49 | No clear effect |
| Add vectree quantization (ablation) | CityGaussianV2 with SH degree 2 · Residence (GauU-Scene) | F1: 0.474 -> 0.465 | Hurt |
| Use 7k-iteration pretraining (ablation) | CityGaussianV2 · Residence (GauU-Scene) | F1: 0.465 -> 0.461 | Hurt |
Limitations
- Evaluated only on scenes with compensated or constant lighting; illumination variation not addressed.
- TSDF-based mesh extraction struggles with thin structures such as spires and occluded road surfaces and building facades.
- Compression still lags CityGS in rendering speed even at similar Gaussian counts.
Experimental setup
Datasets: GauU-Scene MatrixCity
Benchmarks: GauU-Scene MatrixCity
Metrics: PSNR SSIM LPIPS F1 score precision recall training time memory storage size FPS
Baselines: SuGaR 2DGS GOF NeuS Neuralangelo CityGS
Builds on: 2D Gaussian Splatting CityGaussian 3D Gaussian Splatting Depth Anything V2 LightGaussian
Closest papers by meaning
- VastGaussian: Vast 3D Gaussians for Large Scene Reconstruction (CVPR 2024 · similarity 0.92)
- EVolSplat: Efficient Volume-based Gaussian Splatting for Urban View Synthesis (CVPR 2025 · similarity 0.91)
- Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization (CVPR 2025 · similarity 0.90)
- LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and 200+ FPS (NeurIPS 2024 · similarity 0.90)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
- DOGS: Distributed-Oriented Gaussian Splatting for Large-Scale 3D Reconstruction Via Gaussian Consensus (NeurIPS 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.