Abstract:

Traditional 3D Gaussian Splatting (3DGS) relies on time-consuming and fragile Structure-from-Motion (SfM) preprocessing, leading to significant application bottlenecks. Although recent approaches leverage pretrained Multi-View Stereo (MVS) models to bypass SfM, they suffer from high computational cost and memory overhead, rendering them unsuitable for dense-view or large-scale reconstruction. To address these issues, an efficient 3DGS framework based on VGGT preprocessing and adaptive voxelization was proposed, which reduced dependence on input views, camera poses, and training resources, enabling end-to-end scene reconstruction within minutes. Firstly, VGGT required only input images and was used to infer camera poses and initial dense scene point clouds within seconds. Then, a depth-refined point-cloud reconstruction module was designed to enhance geometric completeness, boundary sharpness, and fine-grained realism of the initial point cloud. Then, an adaptive voxelization strategy guided by multi-dimensional image saliency was introduced to prune redundant Gaussian primitives during training, significantly reducing memory usage. Finally, confidence-aware depth regularization was combined with multi-view geometric consistency constraint to compensate for rendering quality degradation after voxelization, achieving an optimal balance between compression efficiency and visual fidelity. Experimental results demonstrate that, compared with the current advanced methods, the proposed method achieves consistent improvements in PSNR, SSIM, and LPIPS on the TNT (sparse and dense views) and Mip-NeRF360 (sparse views) datasets. Meanwhile, the number of Gaussian primitives is reduced by approximately 25% and 67% in sparse and dense scenes, respectively, enabling fast and high-quality scene reconstruction.

Key words: novel view synthesis, 3D Gaussian splatting, sparse view, dense view, large scene reconstruction