MiniGS: Efficient 3D Gaussian Splatting With Full Factors Weighted Pruning For Scene Representation

Abstract 3D Gaussian Splatting has recently emerged as a novel 3D representation method, gaining recognition for its exceptional rendering speed and visual fidelity. However, these advancements come at a cost: this technique exhibits substantial memory demands, with optimized models often containing millions of Gaussian primitives that can exceed 1 GB of storage. We attribute this inefficiency to inherent redundancy in primitive allocation, where overlapping or non-essential elements remain unoptimized in current implementations. In this paper, we propose a novel memory-efficient Gaussian compression method named MiniGS, featuring Full Factors Weighted (FFW) Importance Score Pruning and Variational Densify Threshold (VDT) components. On one hand, we construct several importance score to selectively combine pixel coordinates, Gaussian distances and opacity on the set of Gaussian primitives, and utilize them to prune out redundancy while preserving a small number of highly contributive primitives, achieving an impressive 90% compression ratio. On the other hand, to compensate for the quality loss of pruning Gaussians, we utilize the plug-and-play Variational Densify Threshold (VDT) Component to recover fine details of the 3D scenes, effectively compensating for quality losses with primitives at a low scale. We demonstrate the performance of MiniGS with extensive experimental results on classic dataset and our self-constructed synthetic dataset. For example, our proposed method MiniGS can achieve 29.63 PSNR with 303.271 thousand Gaussian primitives, while the vanilla Gaussian splatting algorithm achieves same-level PSNR (29.73) with one order of magnitude higher of Gaussian primitives on the DeepBlending dataset.