Compact 3D Gaussian Splatting For Dense Visual SLAM

• DOI: 10.48550/arxiv.2403.11247
• Publication Date: 2024-03-17
• Authors: Tianchen Deng; Yaohui Chen; Leyan Zhang; Jianfei Yang; Shenghai Yuan; Danwei Wang; Weidong Chen
• Keywords: FOS: Computer and information sciences; Computer Science - Robotics; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition; Robotics (cs.RO)

Recent work has shown that 3D Gaussian-based SLAM enables high-quality reconstruction, accurate pose estimation, and real-time rendering of scenes. However, these approaches are built on a tremendous number redundant Gaussian ellipsoids, leading to high memory storage costs, slow training speed. To address the limitation, we propose compact Splatting system reduces parameter size ellipsoids. A sliding window-based masking strategy is first proposed reduce Then observe covariance matrix (geometry) most ellipsoids extremely similar, which motivates novel geometry codebook compress geometric attributes, i.e., parameters. Robust estimation achieved by global bundle adjustment method with reprojection loss. Extensive experiments demonstrate our achieves faster speed while maintaining state-of-the-art (SOTA) quality scene representation.