Abstract
We introduce GSVisLoc, a visual localization method designed for 3D GaussianSplatting (3DGS) scene representations. Given a 3DGS model of a scene and aquery image, our goal is to estimate the camera's position and orientation. Weaccomplish this by robustly matching scene features to image features. Scenefeatures are produced by downsampling and encoding the 3D Gaussians while imagefeatures are obtained by encoding image patches. Our algorithm proceeds inthree steps, starting with coarse matching, then fine matching, and finally byapplying pose refinement for an accurate final estimate. Importantly, ourmethod leverages the explicit 3DGS scene representation for visual localizationwithout requiring modifications, retraining, or additional reference images. Weevaluate GSVisLoc on both indoor and outdoor scenes, demonstrating competitivelocalization performance on standard benchmarks while outperforming existing3DGS-based baselines. Moreover, our approach generalizes effectively to novelscenes without additional training.