In recent years, 3D Gaussian splatting (3D GS) excels in novel view synthesis. However, achieving high-precision geometric reconstruction of large-scale outdoor scenes from drone images based on 3D GS remains a challenging task. Most related studies have focused solely on small-scale outdoor scenes. In this paper, we propose a high-quality mesh reconstruction approach for large-scale scenes. This method first introduces a self-expanding scene blocking technique, which divides the large scene into multiple independent blocks, allowing for parallel training and reconstruction of each block, and finally integrates them to create a large scene mesh. In the reconstruction of each block, a novel progressive self-planarized method is proposed, which simultaneously leverages the advantages of 3D Gaussian representations for complex features and planarized Gaussian representations for mesh. During the block training process, both 3D Gaussian and planarized Gaussian can exist within the scene, and allowing for their mutual transformation. To reduce holes in the mesh, we propose a local enhancement method for the point clouds. Additionally, to eliminate inaccurate texture colors, a threshold-based highlight detection and suppression strategy is introduced. Experiments on public and custom datasets show our method ensures high-precision large-scene mesh reconstruction while maintaining superior novel view synthesis.
Chen et al. (Wed,) studied this question.