HRGS: Hierarchical Gaussian Splatting for Memory-Efficient High-Resolution 3D Reconstruction

Changbai Li, Haodong Zhu, Tongfei Chen, Shuo Yang, Shuwei Shao, Wenhao Dong, Hanlin Chen, Juan Zhang, Baochang Zhang

Keywords: 3DGS; 3D surface reconstruction

Abstract: 3D Gaussian Splatting (3DGS) has achieved significant progress in real-time 3D scene reconstruction. However, its application in high-resolution reconstruction scenarios faces severe memory scalability bottlenecks. To address this issue, we propose Hierarchical Gaussian Splatting (HRGS), a memory-efficient framework with hierarchical block-level optimization from coarse to fine. Specifically, we first derive a global, coarse Gaussian representation from low-resolution data; we then partition the scene into multiple blocks and refine each block using high-resolution data. Scene partitioning comprises two steps: Gaussian partitioning and training data partitioning. In Gaussian partitioning, we contract irregular scenes into a normalized, bounded cubic space and employ a uniform grid to evenly distribute computational tasks among blocks; in training data partitioning, we retain only those observations that lie within their corresponding blocks or make significant contributions to the rendering results. By guiding each block’s refinement with the global coarse Gaussian prior, we ensure alignment and seamless fusion of Gaussians across adjacent blocks. To reduce computational resource demands, we introduce an Importance-Driven Gaussian Pruning (IDGP) strategy: during each block’s refinement, we compute an importance score for every Gaussian primitive and remove those with minimal rendering contribution, thereby accelerating convergence and reducing redundant computation and memory overhead. To further enhance surface reconstruction quality, we also incorporate normal priors from a pretrained model. Finally, even under memory-constrained conditions, our method enables high-quality, high-resolution 3D scene reconstruction. Extensive experiments on three public benchmarks demonstrate that our approach achieves state-of-the-art performance in high-resolution novel view synthesis (NVS) and surface reconstruction tasks.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 10549