Research map / 3D vision / Gaussian splatting
GigaGS: 3D Gaussian Based Planar Representation for Large-Scene Surface Reconstruction
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.
Proposes GigaGS, a 3DGS-based method for large-scale aerial scene surface reconstruction using visibility-based parallel partitioning, LoD anchor representation, and multi-view photometric/geometric consistency losses.
Problem
Surface reconstruction of large-scale outdoor scenes is hard due to GPU memory limits, varying levels of geometric detail, and appearance inconsistencies across aerial images; prior 3DGS surface methods target objects or small scenes.
Approach
Combines an octree-style LoD anchor-based 3DGS with flattened Gaussians, a visibility-driven non-overlapping camera partitioning for parallel training, and local depth-normal and multi-view NCC/geometric consistency losses.
Key results
- On the Building scene GigaGS reaches SSIM 0.905 and PSNR 26.69, versus 0.804/23.50 for VastGaussian and 0.480/16.12 for PGSR.
- Among mesh-producing methods, GigaGS outperforms PGSR, SuGaR, NeuS and Neuralangelo on all five scenes in rendering metrics, while several baselines run out of memory on Campus (SuGaR).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GigaGS instead of VastGaussian | Large-scene reconstruction · Building test views | SSIM: 0.804 -> 0.905 | Helped |
| Use GigaGS instead of VastGaussian | Large-scene reconstruction · Rubble test views | PSNR: 26.92 -> 25.10 | Hurt |
| Use GigaGS instead of VastGaussian | Large-scene reconstruction · Campus test views | SSIM: 0.816 -> 0.773 | Hurt |
| Use GigaGS instead of VastGaussian | Large-scene reconstruction · Residence test views | LPIPS: 0.124 -> 0.190 | Hurt |
| Use GigaGS instead of VastGaussian | Large-scene reconstruction · Sci-Art test views | SSIM: 0.885 -> 0.883 | Hurt |
| Add an appearance model (ablation) | GigaGS · Surface reconstruction | surface artifacts | Helped |
| Add flatten regularization without additional geometric supervision (ablation) | GigaGS · Surface reconstruction | model expressiveness | Hurt |
| Add local depth-normal consistency loss (ablation) | GigaGS · Surface reconstruction | surface quality | Helped |
| Add multi-view regularization (ablation) | GigaGS · Surface reconstruction | surface reconstruction performance | Helped |
Surprising findings
- On the Campus and Residence scenes GigaGS has lower PSNR than VastGaussian (22.79 vs 26.00 and 24.34 vs 26.81), so it is not uniformly best in rendering quality.
Limitations
- Mesh quality evaluated qualitatively; surface metrics are not reported quantitatively.
- Rendering quality lower than VastGaussian on some scenes (Campus, Residence).
Experimental setup
Datasets: Mill-19 Building Mill-19 Rubble UrbanScene3D Campus UrbanScene3D Residence UrbanScene3D Sci-Art
Benchmarks: Mill-19 UrbanScene3D
Metrics: PSNR SSIM LPIPS
Baselines: Mega-NeRF VastGaussian PGSR SuGaR NeuS Neuralangelo
Builds on: 3D Gaussian Splatting OctreeGS PGSR VastGaussian 2DGS
Closest papers by meaning
- VA-GS: Enhancing the Geometric Representation of Gaussian Splatting via View Alignment (NeurIPS 2025 · similarity 0.90)
- Holistic Large-Scale Scene Reconstruction via Mixed Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
- Graph-Guided Scene Reconstruction from Images with 3D Gaussian Splatting (ICLR 2025 · similarity 0.89)
- Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization (CVPR 2025 · similarity 0.89)
- GS-2DGS: Geometrically Supervised 2DGS for Reflective Object Reconstruction (CVPR 2025 · similarity 0.89)
- VastGaussian: Vast 3D Gaussians for Large Scene Reconstruction (CVPR 2024 · similarity 0.89)
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.