Stereo Gaussian Splatting with Adaptive Scene Depth Estimation for Semantic Mapping.

Area of Science:

• Robotics and Computer Vision
• 3D Scene Reconstruction
• Simultaneous Localization and Mapping (SLAM)

Background:

• Accurate geometric and semantic understanding in complex environments is crucial for robotics and augmented reality (AR).
• Existing neural implicit representations for SLAM face challenges like high computational costs and online mapping forgetting.
• Current SLAM systems often rely on active depth sensors, limiting their applicability.

Purpose of the Study:

• To propose StereoGS-SLAM, a stereo semantic SLAM framework leveraging 3D Gaussian Splatting (3DGS) for explicit scene representation.
• To enable robust and scale-consistent reconstruction using only passive RGB stereo inputs.
• To enhance real-time performance and stability through adaptive depth estimation and hybrid keyframe selection.

Main Methods:

• Development of StereoGS-SLAM, a stereo semantic SLAM framework utilizing 3D Gaussian Splatting.
• Implementation of an adaptive depth estimation strategy to dynamically refine Gaussian scales.
• Introduction of a hybrid keyframe selection strategy combining motion-aware and random sampling.

Main Results:

• StereoGS-SLAM achieves robust and scale-consistent 3D reconstruction from passive stereo data.
• The system demonstrates competitive localization accuracy and rendering performance.
• Experimental results show stable, real-time optimization and improved keyframe diversity.

Conclusions:

• StereoGS-SLAM offers a viable alternative to existing SLAM systems, particularly those requiring active sensors.
• The proposed adaptive depth estimation and keyframe selection enhance the efficiency and robustness of 3DGS-based SLAM.
• This framework advances the state-of-the-art in semantic SLAM for complex environment mapping.