Splat-SLAM: Globally Optimized RGB-only SLAM with 3D Gaussians

Erik Sandström, Ganlin Zhang, Keisuke Tateno, Michael Oechsle, Michael Niemeyer, Youmin Zhang, Manthan Patel, Luc Van Gool, Martin Oswald, Federico Tombari; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025, pp. 1680-1691

3D Gaussian Splatting offers a compact, efficient approach to RGB-only dense SLAM by providing high-quality map rendering with a dense, optimized 3D Gaussian map. Existing methods, however, often underperform in reconstruction quality compared to alternatives like neural point clouds, primarily due to limited map and pose optimization or reliance on monocular depth. We introduce the first RGB-only SLAM system with globally optimized tracking, dynamically adapting the Gaussian map to keyframe pose and depth updates. To address the lack of geometric priors, we incorporate so called Disparity, Scale and Pose Optimization (DSPO) for bundle adjustment, jointly optimizing pose, depth, and monocular depth scale. Our tests on Replica, TUM-RGBD, and ScanNet confirm this approach achieves superior or comparable tracking, mapping, and rendering accuracy with small map sizes and fast runtimes.