Trimming the Fat: Efficient Compression of 3D Gaussian Splats through Pruning

TL;DR

The paper tackles the scalability challenge of differentiable 3D Gaussian Splatting (3DGS) by introducing Trimming the Fat, a gradient-informed post-hoc pruning method. Starting from a pre-trained 3DGS, it iteratively prunes Gaussians using opacity and gradient signals, guided by a 3D prior to preserve scene fidelity, and then fine-tunes the remainder. The approach achieves up to 4× pruning on its own and up to ~25× compression when combined with end-to-end compression, while delivering up to 600 FPS and maintaining high rendering quality across diverse benchmarks. Ablation studies show the necessity of a 3D prior, the superiority of iterative pruning over one-shot pruning, and the method’s advantages over competing compression schemes, underscoring its potential for edge devices and real-time applications.

Abstract

In recent times, the utilization of 3D models has gained traction, owing to the capacity for end-to-end training initially offered by Neural Radiance Fields and more recently by 3D Gaussian Splatting (3DGS) models. The latter holds a significant advantage by inherently easing rapid convergence during training and offering extensive editability. However, despite rapid advancements, the literature still lives in its infancy regarding the scalability of these models. In this study, we take some initial steps in addressing this gap, showing an approach that enables both the memory and computational scalability of such models. Specifically, we propose "Trimming the fat", a post-hoc gradient-informed iterative pruning technique to eliminate redundant information encoded in the model. Our experimental findings on widely acknowledged benchmarks attest to the effectiveness of our approach, revealing that up to 75% of the Gaussians can be removed while maintaining or even improving upon baseline performance. Our approach achieves around 50$\times$ compression while preserving performance similar to the baseline model, and is able to speed-up computation up to 600 FPS.

Figures (8)

• Figure 1: Vanilla 3DGS-30k vs our novel pruning approach applied with an end-to-end compression technique niedermayr2023compressed.
• Figure 2: Overview of our pruning pipeline. From a pre-trained 3DGS-30k scene, we first iteratively prune it for a fixed number of iterations with subsequent fine-tuning. Then, we conduct further fine-tuning for 20,000 iterations to obtain our final optimized scene.
• Figure 3: The graph depicts the trade-off between performance and size when utilizing Trimming the Fat (gradient-aware iterative pruning) compared to the 3DGS-30k and 3DGS-7k baselines as well as opacity-based pruning on the Mip-NeRF360, Tanks&Temple, and Deep Blending datasets.
• Figure 4: Qualitative comparison of the garden scene at various pruning levels ($\gamma_{\text{iter}}$) using Trimming the Fat (gradient-aware iterative pruning). Our proposed method demonstrates substantially higher compression rates compared to both baselines while maintaining similar visual quality.
• Figure 5: Trade-off between performance and size through iterative pruning and one-shot pruning techniques (a) and in terms of FPS on the Tanks&Temples dataset (b), and opacity distribution before and after pruning for the truck scene (c).