LGSBA: Local Gaussian Splatting Bundle Adjustment for Optimizing 3DGS Rendering Quality

The 3DGS technology is a hot research direction in the current fields of computer vision and robotic perception. The initialization process of the 3DGS system requires obtaining the initial point cloud of the scene and the camera pose to assist in the establishment of the Gaussian map. Most of the research on initialization methods adopt the COLMAP system, which makes the overall time for establishing the Gaussian map relatively long. Based on this, in this study, the ORB-SLAM initialization method is first adopted to obtain the initial point cloud and the camera pose, which shortens the initialization time by 10 times. Secondly, in order to improve the reconstruction quality of the 3DGS, we introduce the LGSBA algorithm during the 3DGS training process. At the same time, our system uses ROS to achieve a tight coupling between the ORB-SLAM and the 3DGS system. On the basis of not affecting the reconstruction quality of the 3DGS, we finally achieve a significant superiority over the system that combines COLMAP and the traditional 3DGS in terms of both optimization time and optimization quality.