Recognition and 6D Pose Estimation of Large-scale Objects using 3D Semi-Global Descriptors

Abstract

While the main focus of 3D object recognition is on small human manipulable objects, we face the problem of recognizing large-scale objects. These objects have a huge impact on mobile robot manipulation and navigation tasks, especially if the object is only partially visible and due to its size is far away from the camera and robot. In our work, we propose a framework capable of recognition and pose estimation for large-scale objects. We propose the use of semi-global descriptors for scene segments and model views in combination with up-sampling and segment label merging techniques. To achieve high accuracy, the initially estimated pose is first refined and afterwards verified. A performance comparison between different model descriptors shows that the chosen semi-global descriptor gives most promising results. By using simultaneous reconstruction, segmentation and recognition, we have built a framework which recognizes large-scale objects and estimates their 6D poses.