Research map / 3D vision / Gaussian splatting
Effective Rank Analysis and Regularization for Enhanced 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.
Proposes effective rank analysis of 3D Gaussian covariance matrices, showing many Gaussians collapse to needle-like shapes (erank≈1), and adds an effective-rank regularizer plus a revised densification rule to improve geometry, normals, and novel view synthesis as an add-on to 3DGS, SuGaR, 2DGS, and GOF.
Problem
3D Gaussian Splatting produces needle-like artifacts, suboptimal geometry and inaccurate normals because Gaussians converge to anisotropic shapes with one dominant variance, which existing flatness-based regularizers do not distinguish from disk-like shapes.
Approach
Uses the differentiable effective rank of each Gaussian's covariance as a shape statistic and as a regularizer that penalizes erank approaching 1 (needle-like) while favoring disk-like (erank≈2) Gaussians, combined with a summed-norm densification criterion that helps split disk-like Gaussians.
Key results
- On DTU, adding the effective rank regularizer lowers mean Chamfer distance for 3DGS from 1.96 to 1.04 (3DGS+e 1.04 vs 3DGS 1.96) and raises PSNR from 32.82 to 33.09.
- SuGaR+e reduces mean Chamfer distance from 1.33 to 1.00, and GOF+e from 0.74 to 0.66 on DTU.
- On Mip-NeRF360, 3DGS+e improves average outdoor PSNR from 24.64 to 24.93 and indoor PSNR from 31.13 to 31.16, while reducing needle count in scene 37 from 16320 to 23 at 30k iterations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add effective rank regularization | 3DGS · DTU | mean Chamfer distance: 1.96 -> 1.04 | Helped |
| Add effective rank regularization | 3DGS · Mip-NeRF360 outdoor scene | PSNR: 24.64 -> 24.93 | Helped |
| Add effective rank regularization | 3DGS · Mip-NeRF360 indoor scene | PSNR: 31.13 -> 31.16 | Helped |
| Add effective rank regularization (ablation) | 3DGS · DTU | mean Chamfer distance: 1.96 -> 1.03 | Helped |
| Add fixed densification (ADC) algorithm (ablation) | 3DGS · DTU | mean Chamfer distance: 1.96 -> 1.26 | Helped |
| Add fixed densification (ADC) algorithm (ablation) | 3DGS · DTU | mean PSNR: 32.82 -> 32.97 | Helped |
| Add effective rank regularization after fixed densification (ablation) | 3DGS + fixed densification · DTU | mean Chamfer distance: 1.26 -> 1.03 | Helped |
| Add effective rank regularization after fixed densification (ablation) | 3DGS + fixed densification · DTU | mean PSNR: 32.97 -> 33.09 | Helped |
| Add optional bag of tricks (ablation) | 3DGS + fixed densification + effective rank regularization · DTU | mean Chamfer distance: 1.03 -> 0.66 | Helped |
| Add optional bag of tricks (ablation) | 3DGS + fixed densification + effective rank regularization · DTU | mean PSNR: 33.09 -> 33.01 | Hurt |
Surprising findings
- Most 3DGS Gaussians with small scales converge to needle-like shapes (erank≈1) rather than disk-like shapes, and the number of needles grows during training while metrics plateau.
- Regularization reduces the number of Gaussians by about 10% and storage, while still improving quality.
Limitations
- The regularizer acts on individual Gaussians and ignores local and global scene structure.
- The hyperparameter lambda_erank is manually selected and may not suit scenes dominated by thin structures.
- Gains on Mip-NeRF360 novel view synthesis are small (e.g., indoor PSNR 31.13 to 31.16).
Experimental setup
Datasets: DTU Mip-NeRF360
Benchmarks: DTU Mip-NeRF360
Metrics: Chamfer distance PSNR SSIM LPIPS storage (MB) training time
Baselines: 3DGS SuGaR 2DGS GOF NeRF BakedSDF Mip-NeRF360 Instant NGP MERF
Builds on: 3D Gaussian Splatting effective rank SuGaR 2D Gaussian Splatting Gaussian Opacity Fields
Closest papers by meaning
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.92)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization (CVPR 2025 · similarity 0.91)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
- 3D Skew Normal Splatting (NeurIPS 2026 · similarity 0.91)
- UVGS: Reimagining Unstructured 3D Gaussian Splatting using UV Mapping (CVPR 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.