Research map / 3D vision / Gaussian splatting
VA-GS: Enhancing the Geometric Representation of Gaussian Splatting via View Alignment
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.
Enhances 3D Gaussian Splatting surface reconstruction by adding edge-aware image loss, edge-weighted normal consistency and smoothing, visibility-aware multi-view photometric alignment, and multi-view deep feature alignment, while preserving novel view synthesis quality.
Problem
3D Gaussian Splatting trained only with image rendering losses yields inaccurate geometry, poor boundary delineation, and inconsistent multi-view alignment, especially under illumination variation.
Approach
Augments the 3DGS rendering objective with edge-aware image gradient supervision, edge-downweighted normal consistency and normal smoothing losses, a visibility- and occlusion-weighted multi-view patch photometric alignment via homographies, and a multi-view pretrained-feature cosine alignment loss.
Key results
- Achieves the lowest mean Chamfer distance on DTU (0.49) versus PGSR (0.53) and 2DGS (0.80).
- Reaches a TNT mean F1-score of 0.54, above PGSR (0.52) and 3DGS (0.09).
- Mip-NeRF 360 average PSNR 27.50, SSIM 0.837, LPIPS 0.174, best LPIPS among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add edge-aware image reconstruction term (ablation) | VA-GS · TNT | F1-score: 0.53 -> 0.54 | Helped |
| add edge-aware weight δ (ablation) | VA-GS normal consistency loss · TNT | F1-score: 0.53 -> 0.54 | Helped |
| add normal consistency loss (ablation) | VA-GS · TNT | F1-score: 0.52 -> 0.54 | Helped |
| add normal smoothing loss (ablation) | VA-GS · TNT | F1-score: 0.51 -> 0.54 | Helped |
| add normal consistency and smoothing losses (ablation) | VA-GS · TNT | F1-score: 0.46 -> 0.54 | Helped |
| add multi-view photometric alignment (ablation) | VA-GS · TNT | F1-score: 0.50 -> 0.54 | Helped |
| add multi-view feature alignment (ablation) | VA-GS · TNT | F1-score: 0.53 -> 0.54 | Helped |
| add photometric and feature alignment (ablation) | VA-GS · TNT | F1-score: 0.36 -> 0.54 | Helped |
| add scale regularization to flatten Gaussian disks (ablation) | VA-GS · TNT | F1-score: 0.54 -> 0.54 | No clear effect |
| increase source views from 1 to 3 (ablation) | VA-GS alignment losses · TNT | F1-score: 0.52 -> 0.54 | Helped |
| increase source views from 2 to 3 (ablation) | VA-GS alignment losses · TNT | F1-score: 0.53 -> 0.54 | Helped |
| increase source views from 3 to 4 (ablation) | VA-GS alignment losses · TNT | F1-score: 0.54 -> 0.54 | No clear effect |
Surprising findings
- Flattening Gaussians with a scale regularization penalty gave no reconstruction gain and degraded novel-view quality on Mip-NeRF 360.
- The feature alignment loss gives only a modest F1 gain on TNT (0.54 vs 0.53 without it) but a larger Chamfer gain on DTU.
Limitations
- Assumes known camera poses as input, which may not be available in real-world or extremely sparse-view settings.
- Training is slower than earlier 3DGS variants (about 20.9 minutes on TNT with all losses).
Experimental setup
Datasets: DTU Tanks and Temples Mip-NeRF 360 Deep Blending
Benchmarks: DTU Tanks and Temples Mip-NeRF 360 Deep Blending
Metrics: Chamfer distance F1-score PSNR SSIM LPIPS
Baselines: 3DGS 2DGS GOF PGSR GS-Pull SuGaR NeuS Neuralangelo RaDe-GS GaussianSurfels GausSurf
Builds on: 3D Gaussian Splatting 2DGS PGSR COLMAP
Closest papers by meaning
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization (CVPR 2025 · similarity 0.91)
- SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering (CVPR 2024 · similarity 0.91)
- Structure Consistent Gaussian Splatting with Matching Prior for Few-shot Novel View Synthesis (NeurIPS 2024 · similarity 0.90)
- MS-GS: Multi-Appearance Sparse-View 3D Gaussian Splatting in the Wild (NeurIPS 2025 · similarity 0.90)
- FatesGS: Fast and Accurate Sparse-View Surface Reconstruction Using Gaussian Splatting with Depth-Feature Consistency (AAAI 2025 · 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.