Gaussian Splatting (GS) was originally designed for realistic rendering of novel views, but its anisotropic 3D Gaussian representation makes it particularly promising for visual localization and SLAM. Recent work as GaussReg has explored loop closure detection via Gaussian registration, improving map consistency and accuracy. However, achieving reliable loop closures remains an open problem, especially in complex environments.
In this paper, we address this issue by generating orthographic bird’s-eye views (BEVs) of optimized Gaussian models. Unlike 3D registration methods, which become inefficient as the number of Gaussians grows, our approach leverages 2D renders thus considerably reducing the registration complexity. Furthermore, orthographic views are free from perspective deformations, so that descriptor invariance to 2D rotation is theoretically sufficient to match points of interest between these views.
Experiments on two public indoor datasets demonstrate that our method achieves higher precision than MapClosure , a previous BEV-based approach, while maintaining comparable speed and performance to GaussReg, a state-of-the-art 3D GS registration method. Code will be made publicly available.