Gaussian Splatting SLAM

Matsuki et al. present MonoGS, the first SLAM system to use 3D GaussianSplatting as its sole scene representation in the difficult monocular setting. Vanilla 3DGS requires accurate per-image camera poses from an offline StructureFromMotion stage; in a SLAM context those poses have to be estimated online from a live monocular stream, with no depth sensor. The paper shows that camera tracking can be done directly against the Gaussian map by differentiating the splat rendering with respect to camera parameters, and that this yields a wide basin of convergence and robust tracking at roughly 3 FPS.

The system is unified: the same Gaussian field supports tracking, mapping, and high-quality NovelViewSynthesis. Tracking minimises a photometric loss between the current frame and a render of the current map from the candidate pose. Mapping grows new Gaussians where the map fails to explain incoming pixels and updates existing ones via the usual splatting optimisation. The authors introduce explicit geometric verification and regularisation steps that exploit the primitive nature of Gaussians to suppress the ambiguities (floaters, sliding surfaces, scale drift) that plague incremental monocular reconstruction. When an RGB-D sensor is available, the same machinery accepts depth as an additional supervision signal without architectural change.

Because primitives are explicit and tied to image evidence, the system is able to reconstruct tiny and even transparent objects that are hard for implicit neural SLAM methods. The paper is broadly contemporary with GS-SLAM by Yan et al., which focuses on RGB-D; together they establish Gaussian-based incremental SLAM as a practical direction and serve as direct successors to Point-SLAM and Co-SLAM. The released MonoGS code is a common reference implementation for monocular Gaussian SLAM.