Research map / 3D vision / Gaussian splatting
3D Student Splatting and Scooping
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 Student Splatting and Scooping (SSS), an unnormalized mixture of learnable Student's t primitives with both positive and negative densities, optimized via SGHMC-based sampling for novel view synthesis.
Problem
3D Gaussian Splatting suffers from limited expressivity and low parameter efficiency (needs many components) for novel view synthesis.
Approach
Replace Gaussian primitives with Student's t distributions with learnable degree of freedom, allow negative-density components ("scooping") to subtract color, and optimize with a principled SGHMC sampler with adaptive friction and noise plus component relocation/recycling.
Key results
- SSS reaches 29.90 PSNR on Mip-NeRF 360 and 24.87 PSNR on Tanks & Temples, the best among compared methods in Table 1 (LPIPS 0.145 and 0.138).
- SSS outperforms all baselines in 15 of 15 reduced-component settings and matches or beats baselines with up to 82% fewer components.
- Ablation and parameter-efficiency experiments indicate negative components reduce the number of components needed (e.g., a torus fit with 2 components vs 5 positive-only).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace 3DGS-MCMC with SSS | 3DGS · Mip-NeRF360 | PSNR: 29.89 -> 29.90 | Helped |
| Replace 3DGS-MCMC with SSS | 3DGS · Mip-NeRF360 | SSIM: 0.900 -> 0.893 | Hurt |
| Replace 3DHGS with SSS | 3DGS · Mip-NeRF360 | LPIPS: 0.178 -> 0.145 | Helped |
| Replace 3DHGS with SSS | 3DGS · Tanks & Temples | PSNR: 24.49 -> 24.87 | Helped |
| Replace 3DGS-MCMC with SSS | 3DGS · Tanks & Temples | SSIM: 0.860 -> 0.873 | Helped |
| Replace 3DHGS with SSS | 3DGS · Tanks & Temples | LPIPS: 0.169 -> 0.138 | Helped |
| Replace Scaffold-GS with SSS | 3DGS · Deep Blending | PSNR: 30.21 -> 30.07 | Hurt |
| Replace Scaffold-GS with SSS | 3DGS · Deep Blending | SSIM: 0.906 -> 0.907 | Helped |
| Replace Fre-GS with SSS | 3DGS · Deep Blending | LPIPS: 0.240 -> 0.247 | Hurt |
| Add a negative component alongside a positive component (ablation) | SSS · Torus shape-topology fitting | component count needed to capture topology | Helped |
| Vary the maximum component number (ablation) | SSS · Mip-NeRF360, Tanks & Temples, Deep Blending | PSNR | Helped |
Surprising findings
- Positive-only mixtures need at least 5 components to capture a torus topology, while one positive plus one negative component suffices.
Limitations
- Primitives restricted to symmetric, smooth t-distributions
- SGHMC sampling requires hyperparameter tuning such as the percentage of negative components
- Baselines not run in identical settings, requiring adaptation
Experimental setup
Datasets: Mip-NeRF 360 Tanks & Temples Deep Blending
Benchmarks: Mip-NeRF 360 Tanks & Temples Deep Blending
Metrics: PSNR SSIM LPIPS
Baselines: Mip-NeRF 3DGS GES 3DHGS Fre-GS Scaffold-GS 3DGS-MCMC
Builds on: 3D Gaussian Splatting 3DGS-MCMC mixture models with negative components Generalized Exponential Splatting
Closest papers by meaning
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.91)
- UVGS: Reimagining Unstructured 3D Gaussian Splatting using UV Mapping (CVPR 2025 · similarity 0.91)
- Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS (NeurIPS 2025 · similarity 0.91)
- LP-3DGS: Learning to Prune 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.91)
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · 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.