A ROS-Based Online System for 3D Gaussian Splatting Optimization: Flexible Frontend Integration and Real-Time...

Area of Science:

• Computer Vision
• Robotics
• Computer Graphics

Background:

• Traditional Structure-from-Motion (SfM) methods for 3D Gaussian Splatting (3DGS) initialization are computationally intensive and suffer from decoupled camera pose optimization and map construction.
• Real-time scene reconstruction using 3DGS offers efficiency but faces initialization bottlenecks.

Purpose of the Study:

• To propose and evaluate an online 3DGS optimization system leveraging ROS for improved efficiency and integration.
• To address the computational and decoupling issues in traditional SfM-based 3DGS initialization.

Main Methods:

• Developed a ROS-based system with a loosely-coupled architecture for real-time data interaction between SfM/SLAM frontends and 3DGS backend.
• Integrated various frontend algorithms (e.g., ORB-SLAM, COLMAP) via ROS middleware.
• Implemented a dynamic sliding-window strategy and a combined L1 and SSIM rendering-quality loss function for online 3DGS map optimization.

Main Results:

• Reduced 3DGS initialization time by 90% compared to traditional COLMAP-3DGS.
• Achieved an average Peak Signal-to-Noise Ratio (PSNR) improvement of 1.9 dB across TUM-RGBD, Tanks and Temples, and KITTI datasets.
• Demonstrated effective online optimization of 3DGS maps with improved rendering quality.

Conclusions:

• The proposed ROS-based online 3DGS optimization system significantly enhances initialization speed and reconstruction quality.
• The loosely-coupled architecture and optimized loss function enable efficient and robust real-time 3D scene reconstruction.
• This system offers a flexible and performant alternative for 3DGS applications requiring rapid and accurate scene mapping.