COB-GS: Clear Object Boundaries in 3DGS Segmentation Based on Boundary-Adaptive Gaussian Splitting

Jiaxin Zhang  Junjun Jiang†  Youyu Chen  Kui Jiang  Xianming Liu 
Harbin Institute of Technology    
zhangjiaxin@stu.hit.edu.cn   jiangjunjun@hit.edu.cn   chenyouyu@stu.hit.edu.cn  
† Corresponding author  
CVPR 2025
Paper arXiv github

What is COB-GS?

3D segmentation based on 3D Gaussian Splatting (3DGS) struggles with accurately delineating object boundaries, as Gaussian primitives often span across object edges due to their inherent volume and the lack of semantic guidance during training. Clear Object Boundaries for 3DGS Segmentation aims to improve segmentation accuracy by clearly delineating blurry boundaries of interwoven Gaussian primitives within the scene.

DashGaussian Animation
DashGaussian Animation
Render
SA3D-GS
SAGD
FlashSplat
COB-GS

Method Overview

Boundary-Adaptive Gaussian Splitting and Boundary-Guided Texture Restoration

Architecture

COB-GS is a 3DGS refinement and segmentation method that that jointly optimizes semantics and appearance to register semantic masks to Gaussian primitives. We introduce mask label as an additional attribute to each Gaussian for segmentation. Moreover, we reveal correlation between the gradient direction of labels and the supervising category at the pixel level, which is a strong discriminator of ambiguous Gaussian primitives on the boundary. Speciffcally, during the mask optimization phase, our approach utilizes gradient statistics of the mask label to identify and split boundary Gaussians, allowing precise alignment with object edges. In the scene optimization phase, we refine the scene texture on the correct boundary structure to maintain visual quality.

Robustness Against Erroneous Masks

Architecture

In fact, the binary masks predicted by the trained 2D vision model exhibit discreteness, which often leads to inaccuracies and inconsistencies in object boundary predictions across different views. By refining tiny boundary Gaussians, we exclusively enhance robustness of our method against inaccurate masks from pre-trained model.

Quantitative Results

Segmentation Evaluation

For segmentation evaluation, we utilize the NVOS dataset, which is derived from the LLFF dataset and provides ground truth masks with precise object edges.
Segmentation Evaluation

Visual Evaluation

Due to the lack of reference images for object segmentation results, we use CLIP-IQA, a no-reference IQA that evaluates how well an image matches the given text prompt. The result shows that our method matches "Clear", "Smooth" and "Complete" better than other comparison methods, highlighting our contribution to boundary optimization.
Visual Evaluation

Quantitative Results

Architecture

BibTeX

@inproceedings{zhang2025cobgs,
  title     = {COB-GS: Clear Object Boundaries in 3DGS Segmentation Based on Boundary-Adaptive Gaussian Splitting},
  author    = {Zhang, Jiaxin and Jiang, Junjun and Chen, Youyu and Jiang, Kui and Liu, Xianming},
  booktitle = {CVPR},
  year      = {2025}
}