GS-ReLoc: A Gaussian-Splatting Relocalization Method for Robust and Accurate Mono Camera Pose Estimation

Relocalization is a critical challenge in visual SLAM and autonomous navigation, where precise initial pose estimation is essential for robust system performance. Common approaches to visual relocalization rely on image retrieval to find the most similar image in a database, followed by 2D-2D feature matching to estimate the query image’s pose. However, these methods heavily depend on feature matching, which can be challenging when there is a significant spatial gap between the retrieved and query images or when local visual features are insufficient to establish reliable correspondences. To address image-retrieval database sparsity and reliability we present GS-ReLoc, a novel method that leverages 3D Gaussian Splat (3DGS) models to augment image retrieval databases. This augmentation increases the likelihood of initializing the rendering-based pose refinement process closer to the ground truth (GT) pose, leading to improved final pose estimates. The method begins by constructing a 3DGS model using Structure-from-Motion (SfM) reconstruction, which serves as the foundation for rendering novel virtual keyframes from novel poses. These keyframes enrich the database with diverse viewpoints through an efficient keyframe pose generation strategy. For a query frame, the algorithm identifies the best-matching database entry using a kd-tree structure, providing an initial pose estimate informed by the augmented database. This improved initial pose estimation strategy reduces the risk of the rendering-based pose refinement process converging to local minima, while also yielding higher final pose accuracy. The proposed method is evaluated on the indoor 7Scenes and outdoor Cambridge Landmark datasets, and achieves state-of-the-art pose estimation accuracy while maintaining robustness and computational efficiency, demonstrating its practical applicability in real-world relocalization scenarios.