High-Fidelity Gaussian Splatting from MVS Clouds: An Iterative Spatial Decomposition Framework

Abstract

3D Gaussian Splatting has become a main technique for fast 3D scene reconstruction and editing, leveraging an efficient and flexible explicit representation for high-fidelity real-time rendering. However, the quality of the point clouds used to initialize Gaussians remains a key factor that limits fine-grained geometry reconstruction. To address this limitation, we propose an Iterative Spatial Decomposition (ISD) framework that bridges dense geometric priors from Multi-View Stereo (MVS) with Gaussian Splatting. ISD mitigates the mismatch between dense MVS point clouds and Gaussian sparsity by iteratively partitioning the scene into voxels of adaptive granularity and performing density-aware point assignment. Building on ISD, we introduce Hierarchical Geometric Prior Sampling (HGPS) to substantially reduce redundancy in MVS point clouds while preserving critical details, thereby providing a more robust geometric foundation for reconstruction. We further develop Hierarchical Geometry-aware Initialization (HGI), which uses a voxel-radius-based adaptive parameter initialization and replaces the iterative KNN-based procedure with batch computation, enabling efficient and robust Gaussian initialization. Additionally, we propose a Hierarchical Geometry-aware Densification (HGD) method. By dynamically identifying over-reconstructed or under-reconstructed regions through voxel constraints, HGD enhances detail reconstruction quality while controlling storage overhead. Extensive experiments on Mip-NeRF360, Tanks & Temples, and Deep Blending demonstrate significant improvements in rendering quality, achieving state-of-the-art LPIPS performance. These results indicate that our approach effectively alleviates deficiencies in the geometric priors of initial point clouds and recovers richer geometric details.