Pocket-SLAM: Rendering-Area-Aware Pruning for Memory-Efficient 3DGS-SLAM

3D Gaussian Splatting (3DGS) has become a powerful tool for Simultaneous Localization and Mapping (SLAM) due to its ability to represent complex geometry and synthesize high-quality views. However, in large-scale outdoor environments like autonomous driving scenarios, the number of Gaussian points grows continuously, leading to prohibitively high peak memory consumption. The authors seek to address this bottleneck by developing a memory-efficient pruning strategy that maintains SLAM performance on resource-constrained edge devices.

The authors propose Pocket-SLAM, which integrates two primary strategies to manage Gaussian density. First, they implement a rendering-area-aware pruning strategy that ranks Gaussians based on their actual contribution to the rendered image, rather than relying on local heuristics like opacity or gradient magnitude. Second, to prevent the information loss that typically occurs when pruning texture-dense regions, they introduce a tile-level budget mechanism. This mechanism uses gradient information from the tracking stage to assign survival budgets to different image tiles, ensuring that texture-rich areas retain enough Gaussians to maintain reconstruction fidelity while texture-sparse areas are pruned more aggressively.

Evaluations on the EuRoC and KITTI datasets demonstrate that Pocket-SLAM effectively balances memory efficiency with mapping accuracy. Compared to existing pruning methods, the authors' approach achieves a 60% reduction in peak memory usage and doubles the frames per second (FPS) without sacrificing localization accuracy (measured by ATE) or rendering quality (measured by PSNR). The results suggest that by shifting the focus from local Gaussian properties to scene-level rendering efficiency, 3DGS-SLAM can be scaled more effectively for real-world autonomous navigation.