Mono-SLAM: Monocular 3D Gaussian Splatting SLAM with Geometric Loss and Multi-view Consistency

Recent advances have demonstrated promising results by integrating 3D Gaussian Splatting (3DGS) with SLAM systems for simultaneous localization and photorealistic 3D reconstruction. However, existing monocular GS-SLAM methods rely only on photometric loss for optimizing. Lacking of geometric priors, these methods often suffer from significant trajectory drift in environments with textureless regions, blur, or varying lighting conditions, adversely affecting reconstruction accuracy. In this paper, we propose Mono-SLAM which use a monocular camera as a sensor, and first introduces normal priors and normal loss to compensate for the shortcomings of monocular cameras lacking geometric priors. Additionally, we incorporate multi-view geometric loss along with photometric loss to ensure multi-view consistency, thereby enhancing global geometric accuracy. Finally, we proposed a complete SLAM system that achieves highly accuracy and quality in trajectory estimation and map reconstruction, while demonstrating strong robustness across diverse scenarios.