Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ROI-GS: Interest-based Local Quality 3D Gaussian Splatting

Quoc-Anh Bui, Gilles Rougeron, Géraldine Morin, Simone Gasparini

TL;DR

ROI-GS tackles the inefficiency of uniform resource allocation in 3D Gaussian Splatting by focusing fidelity on user-specified objects of interest. It introduces an object-aware pipeline with object-guided view selection, independent Object-GS training, and a simple composition step that replaces ROI Gaussians in the global Scene-GS. The main contributions are (i) region-aware reconstruction via object-guided view selection and ROI-centric training, (ii) a model-optimization–driven image selection strategy to maximize ROI coverage, and (iii) a practical composition method to integrate high-fidelity Object-GS into the scene. Experiments demonstrate significant ROI quality gains (up to $2.96$ dB PSNR) and a memory footprint reduction of around $17\%$, with faster training when targeting a single ROI, outperforming existing 3DGS baselines. This approach enables high-detail reconstructions for cultural heritage scenes while maintaining real-time rendering, suggesting strong practical impact for digital twins and interactive visualization.

Abstract

We tackle the challenge of efficiently reconstructing 3D scenes with high detail on objects of interest. Existing 3D Gaussian Splatting (3DGS) methods allocate resources uniformly across the scene, limiting fine detail to Regions Of Interest (ROIs) and leading to inflated model size. We propose ROI-GS, an object-aware framework that enhances local details through object-guided camera selection, targeted Object training, and seamless integration of high-fidelity object of interest reconstructions into the global scene. Our method prioritizes higher resolution details on chosen objects while maintaining real-time performance. Experiments show that ROI-GS significantly improves local quality (up to 2.96 dB PSNR), while reducing overall model size by $\approx 17\%$ of baseline and achieving faster training for a scene with a single object of interest, outperforming existing methods.

ROI-GS: Interest-based Local Quality 3D Gaussian Splatting

TL;DR

ROI-GS tackles the inefficiency of uniform resource allocation in 3D Gaussian Splatting by focusing fidelity on user-specified objects of interest. It introduces an object-aware pipeline with object-guided view selection, independent Object-GS training, and a simple composition step that replaces ROI Gaussians in the global Scene-GS. The main contributions are (i) region-aware reconstruction via object-guided view selection and ROI-centric training, (ii) a model-optimization–driven image selection strategy to maximize ROI coverage, and (iii) a practical composition method to integrate high-fidelity Object-GS into the scene. Experiments demonstrate significant ROI quality gains (up to dB PSNR) and a memory footprint reduction of around , with faster training when targeting a single ROI, outperforming existing 3DGS baselines. This approach enables high-detail reconstructions for cultural heritage scenes while maintaining real-time rendering, suggesting strong practical impact for digital twins and interactive visualization.

Abstract

We tackle the challenge of efficiently reconstructing 3D scenes with high detail on objects of interest. Existing 3D Gaussian Splatting (3DGS) methods allocate resources uniformly across the scene, limiting fine detail to Regions Of Interest (ROIs) and leading to inflated model size. We propose ROI-GS, an object-aware framework that enhances local details through object-guided camera selection, targeted Object training, and seamless integration of high-fidelity object of interest reconstructions into the global scene. Our method prioritizes higher resolution details on chosen objects while maintaining real-time performance. Experiments show that ROI-GS significantly improves local quality (up to 2.96 dB PSNR), while reducing overall model size by of baseline and achieving faster training for a scene with a single object of interest, outperforming existing methods.

Paper Structure

This paper contains 10 sections, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Object-Focused Camera Selection.
  4. GS Training
  5. Scene–Objects Composition with Gaussians.
  6. Experiments
  7. Experimental setup.
  8. Quantitative and qualitative evaluation.
  9. Ablations.
  10. Conclusion and Perspectives

Figures (3)

  • Figure 1: A comparison of our proposed method against the classic 3D Gaussian Splatting baseline shows that our method improved object detail with minimal increase in overall scene memory.
  • Figure 2: An overview of the proposed Region of Interest-Focused Gaussian Splatting (ROI-GS) framework, consisting of two stages: scene decomposition and composition. In decomposition, the scene is divided into Scene and Objects groups, with camera sets automatically selected for each GS training. The trained Scene-GS model is used to initialize the Object-GS models. The composition stage integrates high-detail Object-GSs and the global Scene-GS to produce high-quality real-time renderings with enhanced detail for objects in the ROIs.
  • Figure 3: A visual comparison of our method and the baseline. Blurred area are indicated by red arrows.