StaRGS-SLAM: Stable and Robust Gaussian SLAM via Coverage-Aware Fixed-Topology Initialization

WeiTse Cheng

Keywords: Gaussian Splatting, Topology Mutation, Stationary Mapping

Abstract: Reliable 3D scene modeling requires a mapping process that remains stable when early geometry is incomplete or visually ambiguous. We present StaRGS-SLAM, a Gaussian splatting SLAM system that improves robustness and efficiency through a training-free correspondence-guided Gaussian initialization scheme. Instead of relying on residual-driven densification during optimization, StaRGS-SLAM constructs a geometry-aware prior before iterative updates. From keyframes, it extracts dense correspondences from DINOv3 features, suppresses unreliable matches with confidence-aware inlier filtering, and triangulates the filtered observations to produce a well-covered Gaussian representation in a single initialization stage. This design provides stronger geometric support for early mapping, reduces sensitivity to unstable optimization behavior, and shortens convergence time by about 20%. Experiments on TUM RGB-D and Replica show that StaRGS-SLAM achieves competitive or superior localization and reconstruction performance compared with recent Gaussian-based and point-based SLAM methods, while maintaining real-time mapping throughput of up to 925 FPS.

Submission Number: 6