An efficient large-scale scene reconstruction method based on 3D Gaussian splatting

Three-dimensional (3D) reconstruction based on Gaussian Splatting has achieved remarkable success due to its real-time rendering capabilities. However, for large-scale scene reconstruction, existing methods are still far from satisfactory because of the trade-off between high reconstruction quality and the heavy computational cost. In this study, we propose an efficient large-scale scene reconstruction method based on 3D Gaussian Splatting. Specifically, we replace the global attention mechanism in VGGT with a sparse attention mechanism to improve computational efficiency. Furthermore, we introduce a structural information-guided dropout strategy that adaptively suppresses redundant Gaussians in different depth regions according to their structural information. Extensive experiments on various datasets demonstrate that the proposed method is more memory-efficient than state-of-the-art methods while maintaining high reconstruction accuracy.