OneStep-GSPE: an Efficient 3D Gaussian Splatting Based Image Pose Estimation

Abstract. Accurate image pose estimation within a predefined map is critical for applications such as autonomous driving and urban infrastructure management. Conventional methods predominantly rely on feature correspondences, which often require the presence of specific object categories or involve computationally intensive feature learning processes. Recently, 3D Gaussian Splatting (3DGS) has emerged as a promising scene representation technique, offering high-fidelity novel view synthesis while preserving geometric accuracy. However, existing 3DGS-based pose estimation approaches are mainly tailored to small-scale indoor environments with limited lighting variation. Moreover, they typically rely on iterative optimization, which is computationally demanding and often fails to converge when the initial pose error is significant. This paper introduces OneStep-GSPE, a novel and efficient image pose estimation framework designed for outdoor environments with coarse initial poses. By integrating dense LiDAR priors into the 3DGS pipeline, the accuracy of Gaussian initialization is substantially improved, resulting in enhanced scene geometry reconstruction. Furthermore, rendered depth maps are utilized to lift 2D correspondences into 3D space, establishing 2D-3D matches for absolute pose estimation. The proposed method is category-agnostic and eliminates the need for iterative refinement, enabling fast and precise pose estimation. Experiments conducted on the KITTI-360 dataset demonstrate the effectiveness and robustness. OneStep-GSPE achieves a single-image pose estimation time of approximately 1.81 seconds, yielding over a 90% improvement in computational efficiency compared to the baseline. The project page is publicly available.