Pose Averaging from Relative Translations using a Riemannian Staircase

Localization via Pose Graph Optimization is a relevant task in both robotics and computer vision. Most formulations rely on the availability of relative rotations between nodes. However, this is a non-trivial assumption, as in many practical settings (e.g., vision-based camera systems) estimating relative rotations is hard. State-of-the-art methods for pose averaging such as SE-Sync are based on relaxations and are able to obtain low cost solutions, but require relative rotations. In this paper, we present an SE-Sync-style solution for translation-only pose averaging. Our method is based on an application of the Riemannian staircase, and we analyze the potential symmetries introduced by degenerate cases (nodes with only two translation measurements). In particular, we show that our method outperforms other baseline algorithms (such as block coordinate descent) in terms of precision and accuracy. The improvements are particularly apparent in the case of a random initial guess.