GAD-SLAM: 2D Gaussian splatting SLAM via active densification

This paper presents a representation-decoupled dense SLAM framework based on Gaussian scene representations. We maintain a persistent global 3D Gaussian map for dense scene reconstruction, while performing camera tracking using a view-dependent 2D Gaussian proxy dynamically instantiated from the global map. Benefiting from the explicit surface orientation of 2D Gaussian primitives, the proposed tracking method provides more geometrically interpretable image-space projections and more stable gradients, enabling robust and accurate camera pose estimation. For dense mapping, we introduce an active Gaussian densification and visibility pruning strategy that efficiently refines the spatial distribution of 3D Gaussians. By aggressively increasing the number of informative Gaussians in the early optimization stage and removing invisible primitives, the proposed strategy accelerates convergence while maintaining high reconstruction quality. Experimental results demonstrate that our method achieves robust camera tracking, high-quality dense scene reconstruction while maintaining computational efficiency.