RGS-SLAM: Robust Gaussian Splatting SLAM with One-Shot Dense Initialization

TL;DR

RGS-SLAM replaces the residual-driven densification of GS-SLAM with a training-free, one-shot dense initialization derived from dense multi-view correspondences refined by DINOv3 and a confidence classifier. This seeded topology is then refined through a differentiable 3DGaussian Splatting renderer with analytic SE(3) Jacobians, yielding stationary optimization and faster convergence. The approach improves localization accuracy, rendering fidelity, and mapping throughput while remaining compatible with existing Gaussian-SLAM pipelines, demonstrated on Replica and TUM RGB-D with up to 925 FPS. The work offers a practical and scalable pathway to robust, high-fidelity dense SLAM in texture-rich and cluttered indoor environments, while noting limitations in highly dynamic scenes and the need for calibrated sensing and privacy-conscious deployment.

Abstract

We introduce RGS-SLAM, a robust Gaussian-splatting SLAM framework that replaces the residual-driven densification stage of GS-SLAM with a training-free correspondence-to-Gaussian initialization. Instead of progressively adding Gaussians as residuals reveal missing geometry, RGS-SLAM performs a one-shot triangulation of dense multi-view correspondences derived from DINOv3 descriptors refined through a confidence-aware inlier classifier, generating a well-distributed and structure-aware Gaussian seed prior to optimization. This initialization stabilizes early mapping and accelerates convergence by roughly 20\%, yielding higher rendering fidelity in texture-rich and cluttered scenes while remaining fully compatible with existing GS-SLAM pipelines. Evaluated on the TUM RGB-D and Replica datasets, RGS-SLAM achieves competitive or superior localization and reconstruction accuracy compared with state-of-the-art Gaussian and point-based SLAM systems, sustaining real-time mapping performance at up to 925 FPS.

Figures (9)

• Figure 1: Overview of the proposed RGS-SLAM pipeline. The system integrates dense feature matching and multi-view triangulation for one-shot Gaussian initialization, followed by differentiable 3DGS optimization and real-time tracking.
• Figure 2: Detailed RGS-SLAM pipeline. Each keyframe triggers dense multi-view triangulation that yields a one-shot Gaussian initialization, subsequently refined through joint tracking and mapping within a differentiable 3DGS renderer using analytic SE(3) Jacobians.
• Figure 3: Trajectory comparison on a living-room scene. The red line indicates the ground-truth path and the green line shows the estimated trajectory. Our method aligns more closely with the ground-truth and exhibits fewer large drifts than previous systems.
• Figure 4: Rendering results on the TUM dataset. The proposed keyframe-triggered single-step initialization produces sharper edges, fewer transparency artifacts, and more consistent colors than residual-driven densification.
• Figure 5: Tracking error versus Gaussian count. Error decreases rapidly with denser seeding and plateaus near 1000 Gaussians, indicating diminishing returns beyond this density.