Compact 3D Gaussian Splatting for Dense Visual SLAM

Recent work has shown that 3D Gaussian-based SLAM enables high-qualityreconstruction, accurate pose estimation, and real-time rendering of scenes.However, these approaches are built on a tremendous number of redundant 3DGaussian ellipsoids, leading to high memory and storage costs, and slowtraining speed. To address the limitation, we propose a compact 3D GaussianSplatting SLAM system that reduces the number and the parameter size ofGaussian ellipsoids. A sliding window-based masking strategy is first proposedto reduce the redundant ellipsoids. Then we observe that the covariance matrix(geometry) of most 3D Gaussian ellipsoids are extremely similar, whichmotivates a novel geometry codebook to compress 3D Gaussian geometricattributes, i.e., the parameters. Robust and accurate pose estimation isachieved by a global bundle adjustment method with reprojection loss. Extensiveexperiments demonstrate that our method achieves faster training and renderingspeed while maintaining the state-of-the-art (SOTA) quality of the scenerepresentation.