Li-GS: a fast 3D Gaussian reconstruction method assisted by LiDAR point clouds
Wenzhuo Chen, Ruofei Zhong, Kangfei Wang, Donghai Xie
Abstract
Digital twin technology, serving as a bridge between the physical and digital worlds, has demonstrated significant potential across various fields. In the domain of 3D reconstruction, recent advancements in Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) have introduced novel view synthesis methods for creating digital twins. However, existing methods generally require extended processing times and substantial memory usage to achieve high-quality results. This paper introduces Li-GS, a fast 3D Gaussian Splatting method that leverages LiDAR point clouds to accelerate the training of 3D models with enhanced geometric accuracy. Our method begins with dynamic voxel filtering to downsample distinct regions of the point cloud based on the distribution of image features. The point cloud is then colored for model training, and during this process, we limit the expansion order of spherical harmonics to minimize hardware requirements. Additionally, we propose a data processing pipeline that transforms raw device-collected data into a format compatible with our framework. Experimental evaluations on the ETH3D dataset show that our method outperforms existing approaches in rendering quality, computational efficiency, and resource usage, significantly improving its accessibility for deployment on mainstream hardware systems.