Catadioptric Multiview Pose Estimation for Robotic Pick and Place

Robotic handling of objects requires exact knowledge of the object pose. In this work, we propose a novel vision system, allowing robust and accurate pose estimation of objects, which are grasped and held in unknown pose by an industrial manipulator. For superior robustness, we solely rely on object contour as a visual cue. We address the apparent problems of object symmetry and ambiguous perspective by acquiring multiple views of the object cheaply and accurately, through a mirror system. Self-calibration of the mirror setup allows us to model the mirror geometry and perform metric multiview contour matching with a known 3D model.