Trim 3D Gaussian Splatting for Accurate Geometry Representation

TL;DR

TrimGS addresses the challenge of extracting accurate geometry from explicit 3D Gaussian Splatting by introducing contribution-based Gaussian trimming. It computes per-Gaussian contributions across views, trims low-contribution primitives, and couples this with scale-driven densification and normal regularization to preserve fine geometric details while maintaining rendering quality. The method is shown to improve both mesh- and point-based geometry metrics and to enhance perceptual rendering (LPIPS) with compatible integration for 3DGS and 2DGS baselines. This approach offers a practical pathway to more faithful geometry reconstruction from Gaussian fields, with scalable pruning that prioritizes detail over memory savings alone.

Abstract

In this paper, we introduce Trim 3D Gaussian Splatting (TrimGS) to reconstruct accurate 3D geometry from images. Previous arts for geometry reconstruction from 3D Gaussians mainly focus on exploring strong geometry regularization. Instead, from a fresh perspective, we propose to obtain accurate 3D geometry of a scene by Gaussian trimming, which selectively removes the inaccurate geometry while preserving accurate structures. To achieve this, we analyze the contributions of individual 3D Gaussians and propose a contribution-based trimming strategy to remove the redundant or inaccurate Gaussians. Furthermore, our experimental and theoretical analyses reveal that a relatively small Gaussian scale is a non-negligible factor in representing and optimizing the intricate details. Therefore the proposed TrimGS maintains relatively small Gaussian scales. In addition, TrimGS is also compatible with the effective geometry regularization strategies in previous arts. When combined with the original 3DGS and the state-of-the-art 2DGS, TrimGS consistently yields more accurate geometry and higher perceptual quality. Our project page is https://trimgs.github.io

Figures (6)

• Figure 1: TrimGS exhibits better geometric details and perceptual quality. In (a), TrimGS separates the slender crossbars of a bench. (b) showcases that TrimGS excels in rendering intricate details in both color and normals.
• Figure 2: Illustration of robust normal calculation from rendered depth map.
• Figure 3: Qualitative comparison of meshes on the DTU and MipNeRF360 dataset. Note SuGaR sugar uses a different mesh extraction strategy from others, so it includes some background. 3DGS and SuGaR have messy geometry and 2DGS exhibits slight over-smoothness.
• Figure 4: Comparison of rendering quality (test-set view) between 2DGS and Trim2DGS in MipNeRF360 dataset. Our Trim2DGS enhances perceptual quality in high-frequency regions, mitigating the over-smoothness in 2DGS. Notably, Trim2DGS substantially reduces the storage consumption, credited to our proposed contribution-based trimming technique.
• Figure 5: Visualization of Gaussian centers between 2DGS and Trim2DGS from flowers, bicycle scenes in MipNeRF360 dataset. Trim2DGS exhibits a more uniform Gaussians while significantly reducing storage consumption. A better video illustration can be found in our project page.