Large-scale compact 3D Gaussian splatting SLAM.

Area of Science:

• Computer Vision
• Robotics
• Computer Graphics

Background:

• 3D Gaussian splatting (3DGS) enhances real-time rendering, scene control, and editing in SLAM systems.
• 3DGS balances speed and visual quality but suffers from high memory and storage demands due to numerous Gaussian parameters.
• These demands challenge large-scale applications and resource-constrained hardware like lightweight GPUs.

Purpose of the Study:

• To develop a novel, large-scale, compact 3D Gaussian splatting SLAM framework.
• To minimize memory and storage requirements for 3DGS in SLAM.
• To maintain or improve rendering quality, tracking accuracy, and training efficiency.

Main Methods:

• Implemented dynamic masks and a geometric codebook to eliminate redundant 3D Gaussians.
• Introduced a post-tuning and Gaussian pruning module for optimizing Gaussian representations.
• Evaluated the framework on RGB-D and monocular datasets, including noisy outdoor-like conditions.

Main Results:

• Achieved a five-fold reduction in storage costs compared to state-of-the-art algorithms.
• Demonstrated superior rendering quality and improved tracking accuracy.
• Showcased reduced training time and competitive accuracy in challenging, noisy environments.

Conclusions:

• The proposed compact 3D Gaussian splatting SLAM framework effectively addresses memory and storage limitations.
• The method offers a practical solution for large-scale SLAM applications on resource-constrained hardware.
• The framework exhibits robustness and maintains high performance beyond controlled indoor settings.