GRASP-GS: Geometric Registration and Dual-Stag Saliency Pruning for Efficient 3D Gaussian Splatting

Xinghua Lou, Chuyang Wei, Gezhong Pan, Yanzhi Song, Zhuotao Tian

Keywords: 3D Gaussian Splatting; Real-Time Rendering; Point-Based Modeling;Multi-View & 3D

TL;DR: GRASP-GS achieves compact and high-quality scene representations, enabling efficient real-time rendering while preserving structural integrity

Abstract: 3D Gaussian Splatting (3DGS) has emerged as a powerful paradigm for scene representation, enabling real-time rendering with high visual fidelity by modeling scenes as anisotropic 3D Gaussians. However, existing methods suffer from blurred reconstructions, redundant Gaussians, and high training costs due to sparse Structure-from-Motion (SfM) initialization and heuristic densification. In this paper, we propose GRASP-GS (Geometric Registration and DuAl-Stage Saliency Pruning for 3D Gaussian Splatting), a framework that integrates geometric prior-guided initialization with adaptive saliency pruning. Our method first enhances the initial point cloud by extracting and fusing dense multi-view features with the SfM points through multi-stage refinement. Then, a dual-stage pruning strategy sequentially applies KL-based Rendering Survival Pruning (KL-RSP) to reduce spatial redundancy and Opacity-based Density-Constrained Pruning (ODCP) to eliminate low-contribution Gaussians. Experiments demonstrate that GRASP-GS achieves compact and high-quality scene representations, enabling efficient real-time rendering with enhanced structural integrity and visual quality.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 24058