GLAM-SLAM: Real-time Gaussian Large-scale Mapping via Flow Densification and Spatial Decomposition

As Gaussian Splatting (3DGS) becomes the standard for photorealistic scene reconstruction, applying it to large-scale, long-duration outdoor environments remains difficult. Existing 3DGS-based SLAM systems often struggle with memory constraints, slow processing speeds, or the inability to maintain map consistency over long trajectories. This paper asks how to scale 3DGS for real-time, large-scale mapping without sacrificing the tracking accuracy provided by traditional feature-based SLAM.

The authors propose GLAM-SLAM, a decoupled architecture that separates the tracking and mapping tasks. The system uses the well-established ORB-SLAM2 as a lightweight frontend to handle camera tracking and pose estimation on the CPU. Simultaneously, a dedicated Gaussian mapping module runs on the GPU.

To overcome the limitations of sparse feature-based initialization, the authors introduce two key innovations:

1. Flow-Guided Densification: A module that uses optical flow to establish dense, geometrically consistent correspondences between keyframes. These are filtered using epipolar constraints to provide a dense geometric prior for the Gaussian map, ensuring uniform coverage.
2. Spatial Decomposition: To handle the variability of large-scale outdoor scenes, the system partitions the environment into distinct regions. Each region is assigned its own localized Multi-Layer Perceptron (MLP) to predict Gaussian parameters, preventing the "catastrophic forgetting" that occurs when a single network tries to model an entire, diverse environment.

GLAM-SLAM demonstrates significant improvements in reconstruction quality compared to existing state-of-the-art methods. On challenging, long-sequence datasets like KITTI Odometry and Oxford RobotCar, the system achieves a 15% improvement in reconstruction quality (measured by PSNR) over the second-best performer. The decoupled design ensures that the system maintains real-time performance, effectively balancing the need for precise localization with the computational demands of high-fidelity 3D rendering.