An RGB-D Refinement Solution for Accurate Object Pose Estimation

Digital solutions are more and more employed in the industry with the aim of improving manufacturing processes. In this paper, we consider the issue of accurately localizing objects for augmented reality applications. Augmented reality has become a major asset for manufacturing processes as it enables to interactively provide relevant information about manufactured objects. These applications require a high object localization accuracy to avoid displaying wrong information. However, standard object localization methods do not always meet the industry’s accuracy constraints. We address this problem with a novel RGB-D refinement method in order to provide the optimal object localization accuracy. Given a coarse initial object localization, we iteratively refine the input by combining a geometric constraint and depth consistency. We show that fusing these two constraints enables to significantly improve the object localization accuracy. Our refinement method is evaluated on the challenging Linemod and Occlusion datasets. We demonstrate high accuracy and robustness. Furthermore, we propose quantitative and qualitative results on several industrial objects to show the contribution of our method in the applicative field.