Correspondence-Free Relative Pose Estimation: A Global Approach With Sparse Feature-Guided Directional Embedding

Abstract

Estimating the six degrees of freedom relative poses is a fundamental problem in robotics. Generally, correspondence-based methods are often vulnerable to mismatches in features between the source and the target. This paper presents an alternative approach: estimating the relative pose globally without establishing correspondences. Feature functions derived from Direction3D embeddings and Keypoint Encoder are designed to capture sparse features' rotational and translational information, thereby formulating a correspondence-free optimization problem. In addition, a comprehensive pipeline is built that offers robustness or flexibility to estimate SE(3) or SO(3) transformations. We conduct simulations, ablations, and experiments comparing our method with popular correspondence-based and correspondence-free techniques. The results demonstrate that our approach is robust and outperforms existing methods, potentially marking a potential effort for future research.