SafeguardGS: 3D Gaussian Primitive Pruning While Avoiding Catastrophic Scene Destruction

TL;DR

This paper is the first to categorize 3DGS pruning techniques into two types: Scene-level pruning and Pixel-level pruning, distinguished by their scope for ranking primitives, and discovers that assessing based on color similarity with blending weight is the most effective method for discriminating insignificant primitives.

Abstract

3D Gaussian Splatting (3DGS) has made significant strides in novel view synthesis. However, its suboptimal densification process results in the excessively large number of Gaussian primitives, which impacts frame-per-second and increases memory usage, making it unsuitable for low-end devices. To address this issue, many follow-up studies have proposed various pruning techniques with score functions designed to identify and remove less important primitives. Nonetheless, a comprehensive discussion of their effectiveness and implications across all techniques is missing. In this paper, we are the first to categorize 3DGS pruning techniques into two types: Scene-level pruning and Pixel-level pruning, distinguished by their scope for ranking primitives. Our subsequent experiments reveal that, while scene-level pruning leads to disastrous quality drops under extreme decimation of Gaussian primitives, pixel-level pruning not only sustains relatively high rendering quality with minuscule performance degradation but also provides an inherent boundary of pruning, i.e., a safeguard of Gaussian pruning. Building on this observation, we further propose multiple variations of score functions based on the factors of rendering equations and discover that assessing based on color similarity with blending weight is the most effective method for discriminating insignificant primitives. In our experiments, our SafeguardGS with the optimal score function shows the highest PSNR-per-primitive performance under an extreme pruning setting, retaining only about 10% of the primitives from the original 3DGS scene (i.e., 10x compression ratio). We believe our research provides valuable insights for optimizing 3DGS for future works.

Figures (20)

• Figure 1: An incautious selection of pruning settings can lead to unexpected scene degradation. We reproduce (a)-(d) using the pruning technique from RadSplat Niemeyer2024RadSplat, applied at 20k out of 30k iterations during primitive optimization. The caption of each image indicates importance score threshold (primitives below this are pruned), PSNR, and the number of primitives (Best viewed in zoom).
• Figure 2: Given a set of Gaussian primitives $\mathbf{G}_i$ (a), scene-level pruning (b) evaluates the importance of each primitive across all rays and orders the primitives globally. Because this pruning is applied at the scene level, some rays may end up rendering nothing afterward. In contrast, pixel-level pruning (c) evaluates each primitive's contribution and ranks the primitives accordingly independently for each ray, ensuring that primitives are preserved in a way that completely prevents rays from missing intersections with primitives.
• Figure 3: Averaged PSNR over nine scenes from MipNeRF360 Barron2022Mipnerf360. We sweep pruning parameters for each method. Unlike scene-level pruning, pixel-level pruning maintains the number of primitives above a certain boundary, preventing drastic scene degradation while preserving high quality.
• Figure 4: The score function of EfficientGS ($Score_{EG}$) lacks consideration of the colors of Gaussian primitives, resulting in the loss of high-frequency information during pruning, as indicated by the white holes on the wheel axle in (b). Our proposed score function ($Score_{SG\_36}$), which is sensitive to the color similarity between the primitive color and the ground truth pixel value, helps preserve the details in (c).
• Figure 5: We perform a top-k parameter sweep across a group of proposed score functions, as listed in \ref{['table:1']}. Results are averaged over nine scenes from MipNeRF360 Barron2022Mipnerf360.