Six-DoF Pose Estimation With Efficient 3-D Gaussian Splatting Representation for Visual Relocalization

Abstract

Autonomous systems are being increasingly used throughout the globe as a means for exploration, navigation, inspection, and automatic scene perception. Mobile robots, i.e., the core carrier of autonomous systems, require the efficient map-based relocalization component for real-world application. Given the inevitable slow mapping speed, high memory usage and weak visualization of the traditional map, the 3-D Gaussian Splatting (3DGS) radiance field map avoids these problems essentially and shows enormous potential for visual relocalization. With this in view, this article proposes a vision-only six degrees of freedom (6DoF) pose estimation approach based on the explicit 3DGS map representation. Our approach allows a robot to relocalize from red–green–blue images and an efficient pretrained 3DGS map. We deeply study the relations between 6DoF pose and parameters of 3DGS including position mean, covariance and color. Therefore, we formulate a 6DoF pose optimization pipeline by back-propagating analytic gradients from the 3DGS map to the parameterized camera pose. Moreover, we develop our method for visual relocalization in various scenes, which is not limited by parametric setup. In our experiments, we evaluate on the real-world indoor and outdoor scenes. The experiments show faster mapping speed, lower memory usage and comparable accuracy against various baselines for visual relocalization.