Structure-Aware GS-SLAM With Multi-Gaussian Refinement

Existing GS-SLAM systems often underutilize scene structural information and are prone to Gaussian map structure degradation during training, limiting both tracking robustness and reconstruction quality. We present SAGS-SLAM that fully exploits structural cues across tracking, keyframe management, and mapping. For tracking, we integrate global point cloud geometry with Gaussian covariance to perform structure-aware GICP pose estimation, enabling fast and accurate tracking independent of color cues. For keyframe selection and training, we leverage motion-induced structural changes to select informative keyframes, and introduce a dynamic weight allocation mechanism (DWAM) that yields more balanced keyframe usage, allowing high-quality reconstruction with fewer keyframes. For mapping, we introduce a structure-enhanced mapping strategy (SEMS) framework that strengthens local geometric representation, and employ a transparency regularizer to enforce structural consistency by suppressing redundant Gaussians and accelerating convergence. Experiments demonstrate that SAGS-SLAM achieves superior tracking accuracy and higher reconstruction fidelity than existing GS-SLAM frameworks, with improved PSNR, SSIM, and LPIPS, and superior performance in novel perspective synthesis.