EAGS-SLAM: Edge-Assisted Gaussian Splatting SLAM

Existing 3D Gaussian Splatting (3DGS)-based Simultaneous Localization and Mapping (SLAM) methods improve tracking accuracy and rendering quality by incorporating additional components such as loop closure. However, this improvement comes at the cost of reduced speed, and these methods have limited the use of image information. In this paper, we propose EAGS-SLAM, an edge-assisted RGB-D Gaussian Splatting SLAM that leverages edge information to improve system performance. EAGS-SLAM combines edge-based visual odometry and gaussian tracking, which reduces camera tracking time by up to 50%. We propose a novel gaussian seeding method that directly utilizes edge information to capture geometric details in images. A decoupled submap design that can run loop closure in parallel, ensuring globally consistent mapping while improving system speed. The proposed system is evaluated on real-world datasets Scan-Net, TUM RGB-D, and synthetic dataset Replica. Compared with existing state-of-the-art systems, the proposed system demonstrates competitive or superior performance in tracking, rendering, and runtime. The code and visualization results are available at https://endermands.github.io/EAGSSLAM.