Research map / 3D vision / Gaussian splatting

ReorgGS: Equivalent Distribution Reorganization for 3D Gaussian Splatting

NeurIPS 2026 · 3D vision › Gaussian splatting · approach: Compact and efficient Gaussian splatting · 0 citations (OpenAlex)

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 ReorgGS, which resamples centers from a converged 3DGS model's opacity-weighted Gaussian mixture, re-estimates kNN covariances, sets low opacity, and continues optimization to improve fit at a fixed Gaussian budget.

Problem

Converged 3D Gaussian Splatting models approximate the scene but become poorly parameterized for further optimization: floaters block gradients via alpha compositing and redundant overlapping Gaussians cause parameter coupling, so continued optimization plateaus.

Approach

Treat the converged Gaussian set as an empirical mixture field, resample new centers with opacity-proportional categorical sampling, re-estimate anisotropic covariances via kNN, and initialize low opacity, rebuilding the visibility/overlap graph rather than merely resetting opacity on the old graph.

Key results

• Across five backbones and two LiteGS budget variants, ReorgGS improves training PSNR over direct continuation by about 0.5 dB (e.g., 3DGS 32.61 to 33.16 train PSNR).
• Test PSNR gains over direct continuation are about 0.14 dB on average, e.g., 3DGS 28.72 to 28.89.
• Reported rendering speedup of 12.8% on LiteGS (10.75 s to 9.37 s inference time) with matched Gaussian count.

Surprising findings

• Direct continuation already includes periodic opacity resets, and the paper argues pure opacity reset is insufficient because the optimizer restores opacities on the same overlap graph.
• In the half-budget LiteGS setting, direct continuation slightly beats ReorgGS on LPIPS.

Limitations

• Cannot recover structures absent or never observed in the original Gaussian set.
• Effectiveness may drop when the input distribution has severe reconstruction errors.
• Requires additional post-training optimization after reorganization, increasing total training time.

Experimental setup

Datasets: Mip-NeRF 360 Blender Tanks&Temples

Benchmarks: Mip-NeRF 360 Blender Tanks&Temples

Metrics: PSNR SSIM LPIPS

Baselines: Direct continuation 3DGS GOF Taming-3DGS 3DGS-MCMC LiteGS

Builds on: 3D Gaussian Splatting 3DGS-MCMC opacity reset

Closest papers by meaning

• GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
• Metropolis-Hastings Sampling for 3D Gaussian Reconstruction (NeurIPS 2025 · similarity 0.92)
• Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.91)
• AdpSplit: Error-Driven Adaptive Splitting for Faster Geometry Discovery in 3D Gaussian Splatting (NeurIPS 2026 · similarity 0.91)
• Frequency-Aware Density Control via Reparameterization for High-Quality Rendering of 3D Gaussian Splatting (AAAI 2025 · similarity 0.91)
• Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS (NeurIPS 2025 · similarity 0.91)