Research map / 3D vision / Gaussian splatting
POp-GS: Next Best View in 3D-Gaussian Splatting with P-Optimality
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.
Quantifies information gain in 3D Gaussian Splatting for next-best-view selection by applying P-optimality (optimal experimental design) to a Fisher-information-style covariance, with simple diagonal and block-diagonal Hessian approximations.
Problem
3D-GS does not natively quantify uncertainty or information gain from candidate views, which is needed for next-best-view selection, active perception and keyframe selection in SLAM.
Approach
Reformulate information gain in 3D-GS through optimal experimental design, using P-optimality uncertainty functionals (T-, D-, A-, E-optimality) of the covariance derived from a Gauss-Newton/Hessian approximation, plus a block-diagonal per-ellipsoid Hessian approximation that captures intra-ellipsoid correlation.
Key results
- On Mip-NeRF360 with 10 views, D-Opt. (Block) reaches 18.15 PSNR vs 16.81 for FisherRF and 17.29 for uniform sampling.
- On Blender single view selection with 10 views, T-Opt. (Simple) and D-Opt. (Simple) reach 25.40 and 25.52 PSNR vs 24.59 for FisherRF.
- On Blender keyframe selection, D-Opt. (Block) achieves 24.53 PSNR vs 18.37 for FisherRF.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use D-Optimality with a block-diagonal covariance approximation instead of uniform sampling | 3D-GS view selection · Mip-NeRF360, single view selection, 10 views | PSNR: 17.29 -> 18.15 | Helped |
| Use D-Optimality with a block-diagonal covariance approximation instead of FisherRF | 3D-GS view selection · Blender, single view selection, 10 views | PSNR: 24.59 -> 25.41 | Helped |
| Use D-Optimality with a block-diagonal covariance approximation instead of FisherRF | 3D-GS view selection · Mip-NeRF360, single view selection, 20 views | PSNR: 20.89 -> 21.32 | Helped |
| Use D-Optimality with a block-diagonal covariance approximation instead of FisherRF | 3D-GS view selection · Mip-NeRF360, batch view selection | LPIPS: 0.399 -> 0.378 | Helped |
| Use T-Optimality instead of uniform sampling for keyframe selection | 3D-GS keyframe selection · Blender | PSNR: 23.47 -> 24.90 | Helped |
| Add geometric parameters (location, rotation, and scale) to information quantification (ablation) | D-Opt. (Block) · Blender, single view selection, 10 images | PSNR: 25.36 -> 25.41 | Helped |
| Remove spherical harmonics from information quantification (ablation) | D-Opt. (Block) · Blender, single view selection, 10 images | PSNR: 25.41 -> 25.52 | No clear effect |
| Remove opacity from information quantification (ablation) | D-Opt. (Block) · Blender, single view selection, 10 images | PSNR: 25.41 -> 25.56 | No clear effect |
| Use all parameters instead of removing geometry parameters (ablation) | D-Opt. (Block) · Blender, single view selection, 10 images | SSIM: 0.9074 -> 0.9084 | Helped |
| Use all parameters instead of removing geometry parameters (ablation) | D-Opt. (Block) · Blender, single view selection, 10 images | LPIPS: 0.0793 -> 0.0776 | Helped |
Surprising findings
- Uniform sampling outperforms FisherRF on Mip-NeRF360 single view selection with 10 views.
- A- and E-Optimality perform worse than uniform sampling on several settings.
Limitations
- Simple and block diagonal approximations discard correlation between ellipsoids.
- Block diagonal approximation is memory and compute expensive (2.62 GB, 0.15 Hz on ship scene).
Experimental setup
Datasets: Mip-NeRF360 Blender
Benchmarks: Mip-NeRF360 Blender
Metrics: PSNR SSIM LPIPS
Baselines: FisherRF Uniform Sampling A-Opt. (Simple) E-Opt. (Simple) T-Opt. (Simple) D-Opt. (Simple)
Builds on: FisherRF 3D Gaussian Splatting P-optimality PUP 3D-GS
Closest papers by meaning
- GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting (CVPR 2024 · similarity 0.90)
- Gaussian Splatting SLAM (CVPR 2024 · similarity 0.88)
- Metropolis-Hastings Sampling for 3D Gaussian Reconstruction (NeurIPS 2025 · similarity 0.88)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.87)
- Variational Multi-scale Representation for Estimating Uncertainty in 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.87)
- PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting (CVPR 2025 · similarity 0.87)
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.