RP-SLAM: Real-Time Photorealistic SLAM With Efficient 3D Gaussian Splatting.

Area of Science:

• Computer Vision
• Robotics
• Computer Graphics

Background:

• 3D Gaussian Splatting (3DGS) is a high-quality 3D rendering technique with growing interest in Simultaneous Localization and Mapping (SLAM) systems.
• Existing 3DGS-SLAM methods struggle with Gaussian primitive redundancy, forgetting during optimization, and monocular depth initialization challenges.

Purpose of the Study:

• To develop an efficient and photorealistic 3D Gaussian splatting-based SLAM method (RP-SLAM) for both monocular and RGB-D cameras.
• To overcome limitations of existing 3DGS-SLAM approaches, focusing on mapping accuracy, real-time performance, and model compactness.

Main Methods:

• RP-SLAM decouples camera pose estimation from Gaussian primitive optimization.
• Employs an efficient incremental mapping approach with adaptive sampling and filtering for scene representation.
• Utilizes dynamic window optimization to mitigate forgetting and enhance map consistency.
• Introduces a monocular keyframe initialization method using sparse point clouds for improved primitive accuracy.

Main Results:

• RP-SLAM achieves state-of-the-art map rendering accuracy.
• Demonstrates real-time performance suitable for practical applications.
• Maintains model compactness without sacrificing accuracy.

Conclusions:

• RP-SLAM offers a robust and efficient solution for 3D Gaussian splatting-based SLAM.
• The proposed methods effectively address key challenges in integrating 3DGS with SLAM systems.
• RP-SLAM provides a strong foundation for photorealistic and accurate real-time mapping.