Research map / 3D vision / Gaussian splatting
Frequency-Aware Density Control via Reparameterization for High-Quality Rendering of 3D Gaussian Splatting
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.
Links Gaussian density to scale via a reparameterized absolute scale tied to local density, and adds a frequency-aware density control (dynamic gradient threshold plus SSIM-based scale confidence filter) to improve 3DGS quality with fewer Gaussians.
Problem
3DGS lacks an explicit constraint between Gaussian density and scale, so it often uses improper-scale Gaussians for high-frequency details, leading to accuracy loss and redundant Gaussians in novel view synthesis.
Approach
Reparameterize each Gaussian's scale into an absolute scalar tied by an explicit function to local KNN-based density (s_a proportional to D^{-1/3}), and control densification with a dynamic gradient threshold plus a scale-based SSIM confidence filter, so density tracks frequency content.
Key results
- On Mip-NeRF360, FDS-GS reaches SSIM 0.835, PSNR 27.95, LPIPS 0.180, exceeding reproduced 3DGS (0.813/27.48/0.221).
- On Tanks&Temples, FDS-GS reports LPIPS 0.141 vs 0.181 for reproduced 3DGS, with PSNR 24.27 vs 23.89.
- Full model uses about 2.37M Gaussians vs 3.17M for 3DGS in the ablation table, while improving PSNR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FDS-GS instead of 3DGS | 3DGS · MipNeRF-360 | SSIM: 0.815 -> 0.835 | Helped |
| Use FDS-GS instead of 3DGS | 3DGS · Tanks&Temples | PSNR: 23.14 -> 24.27 | Helped |
| Use FDS-GS instead of 3DGS | 3DGS · Deep Blending | SSIM: 0.903 -> 0.906 | Helped |
| Add density-scale linking (ablation) | 3DGS · MipNeRF360 full set | PSNR: 27.48 -> 27.65 | Helped |
| Add density-scale linking (ablation) | 3DGS · MipNeRF360 full set | Gauss.Num.: 3174k -> 5442k | Hurt |
| Add dynamic densification threshold after density-scale linking (ablation) | 3DGS with D&S linking · MipNeRF360 full set | PSNR: 27.65 -> 27.75 | Helped |
| Add dynamic densification threshold after density-scale linking (ablation) | 3DGS with D&S linking · MipNeRF360 full set | Gauss.Num.: 5442k -> 3770k | Helped |
| Replace the dynamic threshold with a fixed gradient threshold (ablation) | FDS-GS · MipNeRF360 full set | rendering quality | No clear effect |
| Replace the scale-based confidence filter with the original 3DGS filter (ablation) | FDS-GS · MipNeRF360 full set | rendering quality | Hurt |
| Replace the scale-based confidence filter with the original 3DGS filter (ablation) | FDS-GS · MipNeRF360 full set | Gauss.Num. | Hurt |
Surprising findings
- Adding only the scale-density linking increases Gaussian count to 5.44M, showing density control alone does not curb growth.
- Using the original 3DGS filter instead of the scale-based confidence filter makes the Gaussian count skyrocket, degrading quality.
Limitations
- Relies on the SSIM-based confidence filter across several observable views, which adds computation.
- Frequency is measured in image space and may vary across viewpoints, which the authors note as a problem for prior frequency methods rather than fully resolved here.
- Hyperparameters (K=50, theta=1.2, grad_preset, tau_c=0.2) are fixed empirically across scenes.
Experimental setup
Datasets: Mip-NeRF360 Tanks&Temples Deep Blending
Benchmarks: Mip-NeRF360 Tanks&Temples Deep Blending
Metrics: SSIM PSNR LPIPS Gaussian number
Baselines: Instant-NGP Mip-NeRF360 3DGS Scaffold-GS Mip-Splatting 3D-HGS FreGS
Builds on: 3D Gaussian Splatting FreGS Mip-Splatting PatchMatch
Closest papers by meaning
- Steepest Descent Density Control for Compact 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
- LODGE: Level-of-Detail Large-Scale Gaussian Splatting with Efficient Rendering (NeurIPS 2025 · similarity 0.91)
- Metropolis-Hastings Sampling for 3D Gaussian Reconstruction (NeurIPS 2025 · similarity 0.91)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
- ReorgGS: Equivalent Distribution Reorganization for 3D Gaussian Splatting (NeurIPS 2026 · similarity 0.91)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
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.