Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Contour Information Aware 2D Gaussian Splatting for Image Representation

Masaya Takabe, Hiroshi Watanabe, Sujun Hong, Tomohiro Ikai, Zheming Fan, Ryo Ishimoto, Kakeru Sugimoto, Ruri Imichi

TL;DR

The authors tackle edge degradation in heavily compressed 2D Gaussian Splatting by introducing contour information-aware segmentation priors. They constrain Gaussian contributions to segmentation regions, employ a warm-up training strategy, and remove pixel clamping to enable richer colors while preserving sharp boundaries. Across synthetic color charts and the DAVIS dataset, their method yields higher edge-focused reconstruction quality and maintains fast rendering with low memory usage, particularly under few-Gaussian regimes. The approach integrates seamlessly with existing 2DGS rasterizers and demonstrates strong practical potential for efficient image representation and editing.

Abstract

Image representation is a fundamental task in computer vision. Recently, Gaussian Splatting has emerged as an efficient representation framework, and its extension to 2D image representation enables lightweight, yet expressive modeling of visual content. While recent 2D Gaussian Splatting (2DGS) approaches provide compact storage and real-time decoding, they often produce blurry or indistinct boundaries when the number of Gaussians is small due to the lack of contour awareness. In this work, we propose a Contour Information-Aware 2D Gaussian Splatting framework that incorporates object segmentation priors into Gaussian-based image representation. By constraining each Gaussian to a specific segmentation region during rasterization, our method prevents cross-boundary blending and preserves edge structures under high compression. We also introduce a warm-up scheme to stabilize training and improve convergence. Experiments on synthetic color charts and the DAVIS dataset demonstrate that our approach achieves higher reconstruction quality around object edges compared to existing 2DGS methods. The improvement is particularly evident in scenarios with very few Gaussians, while our method still maintains fast rendering and low memory usage.

Contour Information Aware 2D Gaussian Splatting for Image Representation

TL;DR

The authors tackle edge degradation in heavily compressed 2D Gaussian Splatting by introducing contour information-aware segmentation priors. They constrain Gaussian contributions to segmentation regions, employ a warm-up training strategy, and remove pixel clamping to enable richer colors while preserving sharp boundaries. Across synthetic color charts and the DAVIS dataset, their method yields higher edge-focused reconstruction quality and maintains fast rendering with low memory usage, particularly under few-Gaussian regimes. The approach integrates seamlessly with existing 2DGS rasterizers and demonstrates strong practical potential for efficient image representation and editing.

Abstract

Image representation is a fundamental task in computer vision. Recently, Gaussian Splatting has emerged as an efficient representation framework, and its extension to 2D image representation enables lightweight, yet expressive modeling of visual content. While recent 2D Gaussian Splatting (2DGS) approaches provide compact storage and real-time decoding, they often produce blurry or indistinct boundaries when the number of Gaussians is small due to the lack of contour awareness. In this work, we propose a Contour Information-Aware 2D Gaussian Splatting framework that incorporates object segmentation priors into Gaussian-based image representation. By constraining each Gaussian to a specific segmentation region during rasterization, our method prevents cross-boundary blending and preserves edge structures under high compression. We also introduce a warm-up scheme to stabilize training and improve convergence. Experiments on synthetic color charts and the DAVIS dataset demonstrate that our approach achieves higher reconstruction quality around object edges compared to existing 2DGS methods. The improvement is particularly evident in scenarios with very few Gaussians, while our method still maintains fast rendering and low memory usage.
Paper Structure (14 sections, 5 equations, 5 figures, 1 table)

This paper contains 14 sections, 5 equations, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Implicit Neural Representation
  4. Gaussian Splatting for Visual Representation
  5. Proposed Method
  6. Contour Information-Aware 2D Gaussian Splatting
  7. Region Assignment of Gaussians
  8. Clamping-Free Training
  9. Experiments
  10. Experimental Setup
  11. Color Chart Results
  12. Real-world Dataset Evaluation
  13. Ablation Study
  14. Conclusion

Figures (5)

  • Figure 1: Overview of the proposed method. Top: The pipeline of our framework. In addition to the conventional 2DGS pipeline, we feed the target image (to be represented with 2D Gaussians) into a segmentation model to obtain object-wise segmentation masks. These masks are integrated into the Gaussian Rasterizer as Object Domain Information. Bottom: Details inside the Gaussian Rasterizer. Each segmentation mask is represented using only the 2D Gaussians that fall within its region.
  • Figure 2: Quantitative results on the synthetic color charts. From left to right: PSNR and MS-SSIM over the entire image, PSNR and SSIM around edge regions.
  • Figure 3: Qualitative results on the synthetic color chart with high compression, where the number of Gaussians is 20. From left to right: ground truth, baseline with clamping, baseline without clamping, ours with clamping, and ours without clamping.
  • Figure 4: Quantitative results on the DAVIS dataset. From left to right: PSNR and MS-SSIM over the entire image, PSNR and SSIM around edge regions. Here, GaussianImage refers to the original method with clamping, while GaussianImage${}^*$ indicates its variant without clamping.
  • Figure 5: Qualitative results on the DAVIS dataset with high compression, where the number of Gaussians is 1250. From left to right: ground truth, input mask for the rasterizer (object domain information), baseline without clamping, and ours.