Stereo-based vs. monocular 6-DoF pose estimation using point features: A quantitative comparison

Abstract

In the recent past, object recognition and localization based on correspondences of local point features in 2D views has become very popular in the robotics community. For grasping and manipulation with robotic systems, in addition accurate 6-DoF pose estimation of the object of interest is necessary. Now there are two substantially different approaches to computing a 6-DoF pose: monocular and stereo-based. In this paper we show the theoretical and practical drawbacks and limits of monocular approaches based on 2D-3D correspondences. We will then present our stereo-based approach and compare the results to the conventional monocular approach in an experimental evaluation. As will be shown, our stereo-based approach performs superior in terms of robustness and accuracy, with only few additional computational effort.