Refining Gaussian Splatting: A Volumetric Densification Approach

2025-08-07

AI Summary

To address the limitation of geometry-agnostic adaptive density control (ADC) in 3D Gaussian Splatting (3DGS)—which hinders novel-view synthesis quality—this paper proposes an inertia-volume-guided volumetric density control method. Specifically, the inertia volume of each Gaussian ellipsoid is leveraged as a geometric prior to guide densification and pruning, enabling more precise point-cloud refinement. Additionally, the paper systematically compares two initialization strategies—Structure-from-Motion (SfM) versus depth-image matching—and analyzes their impact on reconstruction fidelity. Experiments on the Mip-NeRF 360 dataset demonstrate significant improvements in geometric detail recovery and rendering quality, alongside strong generalization across diverse scenes. The core contribution lies in the first integration of inertia volume into 3DGS density control, establishing a geometry-aware adaptive primitive management paradigm.

Technology Category

• Computer Vision: 3D Computer Vision
• Search and Optimization: Sampling/Simulation-based Search
• Machine Learning: Learning with Manifolds

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Security and Privacy: Large-scale security measurements

Abstract

Achieving high-quality novel view synthesis in 3D Gaussian Splatting (3DGS) often depends on effective point primitive management. The underlying Adaptive Density Control (ADC) process addresses this issue by automating densification and pruning. Yet, the vanilla 3DGS densification strategy shows key shortcomings. To address this issue, in this paper we introduce a novel density control method, which exploits the volumes of inertia associated to each Gaussian function to guide the refinement process. Furthermore, we study the effect of both traditional Structure from Motion (SfM) and Deep Image Matching (DIM) methods for point cloud initialization. Extensive experimental evaluations on the Mip-NeRF 360 dataset demonstrate that our approach surpasses 3DGS in reconstruction quality, delivering encouraging performance across diverse scenes.

Problem

• Improving point primitive management in 3D Gaussian Splatting
• Enhancing densification strategy using volumes of inertia
• Evaluating SfM and DIM for point cloud initialization

Innovation

• Volumetric densification using inertia volumes
• Combines SfM and DIM for initialization
• Outperforms 3DGS in reconstruction quality