AMG-SLAM: Adaptive Monocular Gaussian SLAM for Efficient Surface Reconstruction
Youqi Pan, Wugen Zhou, Hongbin Zha
AI summary
Problem
Monocular dense SLAM faces high computational costs, limited reconstruction accuracy, and reliance on resource-intensive dense depth estimation or offline post-processing. Existing Gaussian-based methods lack scene-adaptive awareness, causing redundant Gaussian proposals and uneven optimization across the scene.
Approach
The system tightly couples a learning-based sparse tracking front-end with a Gaussian mapping back-end. It uses a fidelity-aware strategy to propose new Gaussians only in under-reconstructed regions, and a focus-and-balance refinement strategy to prioritize under-optimized Gaussians while maintaining global scene coverage.
Key results
- Fidelity-aware adaptive Gaussian proposal reduces redundancy by targeting under-reconstructed regions
- Focus-and-balance online refinement optimizes high-priority Gaussians while preserving global consistency
- Trajectory, rendering, and geometric accuracy match or exceed state-of-the-art methods
- Fully online, efficient surface reconstruction without dense depth maps or offline post-processing
Why it matters
Provides a computationally efficient, real-time mapping solution for robotics and augmented reality applications that require immediate, high-fidelity 3D environmental understanding.
Abstract
Dense SLAM with a monocular camera remains a highly challenging task. In this paper, we present AMG-SLAM, a novel dense monocular SLAM system that tightly couples sparse tracking with dense Gaussian mapping to achieve fully online and high-quality surface reconstruction. In the front- end, learning-based modules enable efficient pose tracking and Gaussian proposal with sparse depth initialization. Specifically, we propose a fidelity-aware Gaussian proposal strategy that adaptively adds new Gaussians based on reconstruction com- pleteness, effectively avoiding redundancy. In the back-end, we propose a focus-and-balance online refinement strategy, which adaptively selects under-optimized Gaussians for focused refinement while ensuring globally balanced optimization by maximizing scene view coverage. We evaluated our method on synthetic and real-world datasets, including Replica, ScanNet, and EuRoC. Thanks to efficient system coupling and adaptive Gaussian proposal and refinement, our system achieves tra- jectory accuracy, rendering precision, and geometric accuracy comparable to or exceeding current state-of-the-art methods, while also demonstrating high efficiency.