MeGS-SLAM:Memory Efficient Gaussian Splatting SLAM with Graph Signal Processing

Abstract

The recent 3D Gaussian Splatting simultaneous localization and mapping (3DGS-SLAM) has achieved high-fidelity reconstruction from RGB-D images. However, the input lacks edge information in the scene, resulting in inability of Gaussians to accurately model edges and appearance of artifacts at object edges. 3DGS-SLAM involves in a point-based representation method. It produces massive Gaussians for obtaining a detailed map, causing low rendering speed and high storage. To overcome above shortages, we propose a novel SLAM frame with edge priors constraint that adds edge attributes to Gaussians for expressing edge information and edge loss is further introduced to enable Gaussians to accurately reconstruct edges and suppress artifacts. Furthermore, we propose graph signal processing for local Gaussians to establish relationships among irregular Gaussians and extract geometric features from Gaussian scene representations, which are used to efficiently reduce redundant Gaussians without sacrificing performance of tracking and reconstruction. Experiments performed on synthetic and real-world datasets show that our method achieves over 2× compression in memory usage and increases nearly 250% rendering speed while maintaining tracking and mapping performance. Additional information can be found on our project page: yoona12.github.io/MeGS-SLAM.github.io