GaussianTrimmer: Online Trimming Boundaries for 3DGS Segmentation

Liwei Liao, Ronggang Wang · 2026

Read GaussianTrimmer: Online Trimming Boundaries for 3DGS Segmentation by Liwei Liao, Ronggang Wang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-12683.txt.

Introduction

Liwei Liao and Ronggang Wang work on segmenting scenes built from 3D Gaussians. Such scenes represent a place as a cloud of soft Gaussian blobs. Segmentation methods for these scenes have emerged as the format spread. Most of them cut the scene at the level of single Gaussian primitives.

That primitive level cut has a flaw at object edges. Gaussians vary a lot in scale, and large ones often span foreground and background. Cutting by primitive then yields jagged boundaries on segmented objects. Their method is called GaussianTrimmer, and it trims boundaries online. Cleaner boundaries matter for editing, robotics, and virtual scene work. The work targets a visible artifact that users notice at once.

Catalog record

CREATORLiwei Liao, Ronggang Wang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE21 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-12683

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

GaussianTrimmer: Online Trimming Boundaries for 3DGS Segmentation Liwei Liao, Ronggang Wang 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. I Introduction In recent years, 3D Gaussian splatting (3DGS) [ 8 ] has gradually occupied an important position in 3D scene representation due to its high-quality rendering and real-time rendering speed. With the widespread application of 3D Gaussians, 3D scene segmentation methods [ 21 , 2 , 4 , 18 , 26 , 7 , 15 , 20 , 11 , 23 , 13 , 24 , 12 ] based on 3D Gaussians have also gradually emerged. However, most existing 3DGS-based segmentation methods suffer from two key limitations: (1) they treat individual Gaussian primitives as the atomic unit for segmentation, and (2) boundary-straddling Gaussians always lie on object boundaries (as illustrated in Fig. 1 (a)). These straddling Gaussians, which overlap both foreground and background regions, hinder the attainment of sharp segmentation boundaries, often resulting in jagged boundaries. Such irregularities not only degrade quantitative metrics like mIoU and mAcc but also compromise visual quality. These shortcomings severely constrain the utility of 3DGS in downstream tasks, including 3D scene editing (e.g., object relocation or compositing) and robotic applications (e.g., grasping simulations). The jagged boundaries of the segmented objects are refered to as “ boundary challenge ”. Fig. 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. To resolve the boundary challenge, we tackle it from the perspective of decomposing straddling primitives, by further splitting each straddling Gaussian into two smaller Gaussians that respectively belong to the object and the background, thereby achieving training-free trimming of 3D Gaussians. Inspired by SAGD [ 7 ] , we adapt an online method for splitting straddling Gaussians, by decomposing the 2D covariance of the projected Gaussians using segmentation boundaries from the 2D view to guide the decomposition ratio of the 3D covariance, and separately computing the new positions corresponding to the decomposed Gaussians. Unlike previous methods, we treat boundary trimming as a form of post-processing and refinement of the segmentation results (as shown in Fig. 1 (b)). To this end, we devise a boundary trimming strategy based on virtual camera planning, thereby enabling segmentation refinement with extremely low latency (approximately 1s). Our key contributions can be summarized as: • We propose GaussianTrimmer, an online boundary trimming method for 3DGS segmentation, which effectively addresses the boundary challenge within approximately 1 second latency. • We design an effective Virtual Camera Planning (VCP) module for more precise 3D Gaussian decomposition through the 2D masks. • Extensive experiments demonstrate our GaussianTrimmer can work as a plug-and-play module to improve the quantitative and qualitative results. II Related Works II-A 3D Neural Scene Segmentation Since 3DGS was established as both training-efficient and capable of real-time rendering, a range of exceptional 3DGS-specific segmentation methods have emerged [ 21 , 2 , 4 , 18 , 26 , 7 , 15 , 20 , 11 , 23 , 13 , 24 , 12 ] . Specifically, GaussianGrouping [ 21 ] links masks across views using a tracker, then applies joint learning to optimize reconstruction and identity encodings simultaneously. ClickGaussian [ 4 ] , a post-processing approach, enhances pre-trained 3D Gaussians with two-level granularity features and introduces Global Feature-guided Learning to mitigate mask inconsistencies across views. Certain studies [ 7 , 23 ] focus on refining boundaries in 3DGS segmentation. Meanwhile, Flashsplat [ 18 ] accelerates the optimization process by approximately 50 times, though it still requires 30 seconds for optimization, classifying it as an offline method. II-B Boundary Refinement Decomposing 3D primitives has been explored in various contexts. In 3D reconstruction, methods like SAGD [ 7 ] decompose Gaussians based on 2D segmentation boundaries to improve object delineation. Other works have focused on decomposing meshes or point clouds to enhance segmentation quality. Our approach builds upon these ideas by introducing an online boundary trimming method specifically designed for 3DGS segmentation, addressing the challenges posed by straddling Gaussians. COB-GS [ 23 ] also addresses boundary issues but focuses on a different aspect of Gaussian representation. In contrast, our method emphasizes efficient and effective trimming of boundaries as a post-processing step, enhancing existing segmentation results with minimal latency. Unlike previous methods that require retraining or complex optimization, our GaussianTrimmer operates as a plug-and-play solution, making it easy to integrate into existing 3DGS segmentation pipelines. Moreover, our method focuses on real-time performance, ensuring that boundary refinement can be achieved with minimal computational overhead. III Method In this section, we introduce GaussianTrimmer, a method for online trimming boundaries in 3DGS segmentation. As illustrated in Fig. 2 , GaussianTrimmer is a post-processing module designed to refine the segmentation results from existing 3DGS segmentation methods. The input to GaussianTrimmer is a jagged segmentation result, and the output result has smoother boundaries. This step can be formulized as follows: \hat{\mathbf{\Theta}_{\mathcal{O}}}=\texttt{GauTrimmer}(\mathbf{\Theta}_{\mathcal{O}};\mathbf{\Theta},\vec{U}), (1) where \mathbf{\Theta}\in\mathbb{R}^{N\times C} represents total Gaussians of the scene, \mathbf{\Theta}_{\mathcal{O}}\in\mathbb{R}^{K\times C} is a subset of \mathbf{\Theta} , denoting the jagged coarse segmentation result obtained from an existing 3DGS segmentation method, and \vec{U} denotes the up vector used for virtual camera orientation. The output \hat{\mathbf{\Theta}_{\mathcal{O}}}\in\mathbb{R}^{K\times C} is the refined segmentation result with improved boundary quality. Note that \mathbf{\Theta} is optional and can be used for background augmentation during virtual view segmentation. Fig. 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. III-A Virtual Camera Planning (VCP) Corresponding to Fig. 2 (A), the first step of our pipeline is Virtual Camera Planning (VCP). Different from previous boundary refinement methods that utilize 2D masks from training views as cues, we employ virtual views that offer better alignment and more comprehensive coverage (see Fig. 3 ). The VCP module generates a set of virtual cameras \mathbf{c}=\{c_{0},c_{1},...,c_{n-1}\} that uniformly cover the segmented object \mathbf{\Theta}_{\mathcal{O}} . To achieve this, we adopt a spherical coordinate system as follows: First, we determine the centroid of the object, which serves as the center of the spherical coordinate system. The centroid L is computed using the positions of the Gaussians in the segmented object \mathbf{\Theta}_{\mathcal{O}} as L=\frac{1}{|\boldsymbol{p}|}\sum_{\mathbf{x}\in\boldsymbol{p}}\mathbf{x}, (2) where \boldsymbol{p}=\{\mathbf{x}_{i}|\mathbf{x}_{i}\in\mathbf{\Theta}_{\mathcal{O}}\} represents the set of

Read the complete text