Research map / 3D vision / Gaussian splatting
Improving Gaussian Splatting with Localized Points Management
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 Localized Point Management (LPM), a plug-in that uses cross-view matched rendering-error zones to add points and reset opacity of occluding Gaussians in 3DGS and SpacetimeGS.
Problem
Adaptive Density Control in 3D Gaussian Splatting uses global gradient and opacity thresholds, which miss under-optimized regions (e.g., transparent areas) and fail to correct ill-conditioned high-opacity points that occlude valid geometry.
Approach
Localize error-contributing 3D zones by matching features (LightGlue) between the current and a neighboring view, back-projecting error regions as cone intersections, then densify points inside those zones and reset opacity of front-most points there, followed by density-aware pruning.
Key results
- 3DGS* + LPM reaches 23.83 PSNR on Tanks&Temples vs 23.67 for retrained 3DGS*, and 29.76 on Deep Blending vs 29.55.
- STGS* + LPM improves Neural 3D Video average PSNR from 31.99 to 32.40 while keeping near-real-time speed (~105 FPS on Flame Salmon subset comparisons).
- On PlayRoom, 3DGS + LPM matches PixelGS quality (30.22 vs 30.09 PSNR) with 186k Gaussians and 23 min training versus 35 min for PixelGS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Localized Point Management | 3DGS* · Mip-NeRF 360 | PSNR: 27.47 -> 27.59 | Helped |
| add Localized Point Management | 3DGS* · Tanks&Temples | PSNR: 23.67 -> 23.83 | Helped |
| add Localized Point Management | 3DGS* · Deep Blending | PSNR: 29.55 -> 29.76 | Helped |
| add Localized Point Management | STGS* · Neural 3D Video (Flame Salmon) | PSNR: 29.48 -> 29.84 | Helped |
| add Localized Point Management | STGS* · Neural 3D Video dataset | PSNR: 31.99 -> 32.40 | Helped |
| add point addition (ablation) | LPM · PlayRoom | PSNR: 30.10 -> 30.22 | Helped |
| add point addition (ablation) | LPM · Truck | PSNR: 25.43 -> 25.61 | Helped |
| add opacity reset (ablation) | LPM · PlayRoom | PSNR: 30.07 -> 30.22 | Helped |
| add opacity reset (ablation) | LPM · Truck | PSNR: 25.52 -> 25.61 | Helped |
Surprising findings
- Simply lowering the densification threshold in 3DGS inflates Gaussian counts (232k to 523k on PlayRoom) and reduces PSNR, so naive densification is counterproductive.
Limitations
- Still follows 3DGS densification rules, so may not be optimal for under-optimized points.
- Relies on feature matching between neighboring views to localize error zones.
- Gains on some scenes are modest (e.g., 0.01-0.02 SSIM/PSNR changes on several datasets).
Experimental setup
Datasets: Mip-NeRF 360 Tanks&Temples Deep Blending Neural 3D Video Dataset
Benchmarks: Mip-NeRF 360 Tanks&Temples Deep Blending Neural 3D Video Dataset
Metrics: PSNR SSIM LPIPS FPS
Baselines: 3DGS 2DGS MipGS PixelGS SpacetimeGS GaussianPro Mip-NeRF 360 iNGP Plenoxels DyNeRF
Builds on: 3D Gaussian Splatting LightGlue SpacetimeGS Adaptive Density Control
Closest papers by meaning
- Steepest Descent Density Control for Compact 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
- ReorgGS: Equivalent Distribution Reorganization for 3D Gaussian Splatting (NeurIPS 2026 · similarity 0.90)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.90)
- SplatFormer: Point Transformer for Robust 3D Gaussian Splatting (ICLR 2025 · similarity 0.90)
- AdpSplit: Error-Driven Adaptive Splitting for Faster Geometry Discovery in 3D Gaussian Splatting (NeurIPS 2026 · similarity 0.90)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 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.