Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GaussianTrimmer: Online Trimming Boundaries for 3DGS Segmentation

Liwei Liao, Ronggang Wang

TL;DR

GaussianTrimmer addresses boundary jaggedness in 3D Gaussian Splatting segmentation by introducing an online post-processing pipeline that trims straddling Gaussians. It combines Virtual Camera Planning to achieve uniform object coverage, Virtual View Segmentation with 2D masks from SAM2, and Boundary Gaussian Decomposition to split problematic Gaussians along their principal axes, guided by 2D segmentation across views. The method is plug-and-play and demonstrates consistent quantitative gains (mIoU, mAcc) and qualitative boundary improvements across multiple baselines and datasets, with latency around 1 second. This approach enables more precise 3D scene editing and robotics applications by delivering sharper object boundaries without retraining or heavy optimization.

Abstract

With the widespread application of 3D Gaussians in 3D scene representation, 3D scene segmentation methods based on 3D Gaussians have also gradually emerged. However, existing 3D Gaussian segmentation methods basically segment on the basis of Gaussian primitives. Due to the large variation range of the scale of 3D Gaussians, large-sized Gaussians that often span the foreground and background lead to jagged boundaries of segmented objects. To this end, we propose an online boundary trimming method, GaussianTrimmer, which is an efficient and plug-and-play post-processing method capable of trimming coarse boundaries for existing 3D Gaussian segmentation methods. Our method consists of two core steps: 1. Generating uniformly and well-covered virtual cameras; 2. Trimming Gaussian at the primitive level based on 2D segmentation results on virtual cameras. Extensive quantitative and qualitative experiments demonstrate that our method can improve the segmentation quality of existing 3D Gaussian segmentation methods as a plug-and-play method.

GaussianTrimmer: Online Trimming Boundaries for 3DGS Segmentation

TL;DR

GaussianTrimmer addresses boundary jaggedness in 3D Gaussian Splatting segmentation by introducing an online post-processing pipeline that trims straddling Gaussians. It combines Virtual Camera Planning to achieve uniform object coverage, Virtual View Segmentation with 2D masks from SAM2, and Boundary Gaussian Decomposition to split problematic Gaussians along their principal axes, guided by 2D segmentation across views. The method is plug-and-play and demonstrates consistent quantitative gains (mIoU, mAcc) and qualitative boundary improvements across multiple baselines and datasets, with latency around 1 second. This approach enables more precise 3D scene editing and robotics applications by delivering sharper object boundaries without retraining or heavy optimization.

Abstract

With the widespread application of 3D Gaussians in 3D scene representation, 3D scene segmentation methods based on 3D Gaussians have also gradually emerged. However, existing 3D Gaussian segmentation methods basically segment on the basis of Gaussian primitives. Due to the large variation range of the scale of 3D Gaussians, large-sized Gaussians that often span the foreground and background lead to jagged boundaries of segmented objects. To this end, we propose an online boundary trimming method, GaussianTrimmer, which is an efficient and plug-and-play post-processing method capable of trimming coarse boundaries for existing 3D Gaussian segmentation methods. Our method consists of two core steps: 1. Generating uniformly and well-covered virtual cameras; 2. Trimming Gaussian at the primitive level based on 2D segmentation results on virtual cameras. Extensive quantitative and qualitative experiments demonstrate that our method can improve the segmentation quality of existing 3D Gaussian segmentation methods as a plug-and-play method.
Paper Structure (14 sections, 9 equations, 5 figures, 4 tables)

This paper contains 14 sections, 9 equations, 5 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. 3D Neural Scene Segmentation
  4. Boundary Refinement
  5. Method
  6. Virtual Camera Planning (VCP)
  7. Virtual View Segmentation (VVS)
  8. Boundary Gaussian Decomposition (BGD)
  9. Experiments
  10. Experimental Setup
  11. Quantitative Results
  12. Qualitative Results
  13. Ablation Study
  14. Conclusion

Figures (5)

  • Figure 1: (a) Straddling Gaussians make it difficult to achieve precise segmentation boundaries; (b) Our GaussianTrimmer effectively trims these boundaries for improved segmentation quality within only approximately 1 second latency.
  • Figure 2: GaussianTrimmer Pipeline : Guided by user interaction, Virtual Camera Generation (VCG) module generates smooth and object-center virtual views, followed by Rendering-Tracking-Pruning (RTP) loop on the generated virtual views to identify Gaussians belonging to the target, which are represented as a Gaussian-level Boolean 3D mask.
  • Figure 3: Our virtual cameras vs. real cameras. Our planned virtual cameras (top row) provide better object coverage and alignment compared to real cameras (bottom row), leading to more accurate trimming results.
  • Figure 4: (a) Illustration of pixel-to-Gaussian mapping; (b) Ilustration of boundary Gaussian detection.
  • Figure 5: Qualitative Results. This figure showcases the direct improvement effects of GaussianTrimmer on the segmentation results of existing methods. From a visual perspective, GaussianTrimmer effectively refines jagged edges and enhances the clarity of segmentation boundaries.