Catadioptric Silhouette-Based Pose Estimation from Learned Models

The automated handling of objects requires the estimation of object position and rotation with respect to an actuator. We propose a system for silhouette-based pose estimation, which can be applied to a variety of objects, including untextured and slightly transparent objects. Pose estimation inevitably relies on previous knowledge of the object’s 3D geometry. In contrast to traditional view-based approaches our system creates the required 3D model solely from the object silhouettes and abandons the need to obtain a model beforehand. It is sufficient to rotate the object in front of the catadioptric camera system. Experimental results show that the pose estimation accuracy drops only slightly compared to a highly accurate input model. The whole system utilizes the parallel processing power of graphics cards, to deliver an auto calibration in 20 s and reconstructions and pose estimations in 200 ms.