FIGS-SLAM: Gaussian splatting SLAM with dynamic frequency control and influence-based pruning

Abstract

Recent advancements in 3D Gaussian Splatting (3DGS)-based SLAM have significantly improved scene reconstruction compared to traditional SLAM and Neural Radiance Field (NeRF) SLAM, enabling high-quality reconstruction, precise pose estimation, and real-time scene rendering. However, existing approaches still suffer from limitations in detail representation, resulting in mediocre rendering quality. To address these challenges, we propose a new SLAM system, FIGS-SLAM, which utilizes a dynamic frequency-controlled coarse-to-detail map building strategy. Initially, low-frequency components are used to establish a coarse map representation, avoiding premature reliance on high-frequency details and effectively mitigating ambiguities introduced by noise and inaccurate data. As the map construction advances, high-frequency components are gradually incorporated, allowing the system to capture and render intricate geometric details accurately. This coarse-to-fine mapping approach progressively refines the scene representation to achieve higher detail and accuracy. Considering that frequency regularization may generate a large number of redundant 3D Gaussian ellipsoids, we further introduce an influence-based Gaussian pruning strategy. This strategy dynamically prunes Gaussian points with minimal impact on the map by evaluating their transparency, transmission rates, and consensus relationships, thereby enhancing the system’s overall accuracy and efficiency. Experimental results on two datasets show that our method improves PSNR, SSIM, and LPIPS metrics by 4% to 11% compared to existing methods, while achieving an effective balance between map memory management, rendering quality, and runtime efficiency.