Observability Properties of Object Pose Estimation

Abstract

Estimating the pose (position and orientation) of 3D objects solely from images is a challenging task of current relevance to a wide range of robotic and artificial intelligence applications. In this paper, we employ the empirical local observability Gramian as a metric for assessing the quality of image-based pose estimation. We show how the Gramian can be used to describe the local “estimatability” of images of static and dynamic objects. We also show connections between the Gramian matrix and symmetries and near-symmetries of an object, and use this result to characterize the subspace of the pose vector which is unobservable using image data. Results are demonstrated using images constructed in a 3D virtual environment.