Dual Branch PnP Based Network for Monocular 6D Pose Estimation

Monocular 6D pose estimation is a functional task in the field of computer vision and robotics. In recent years, 2D-3D correspondence-based methods have achieved improved performance multiview depth data-based scenes. However, for monocular estimation, these are affected by prediction results correspondences robustness perspective-n-point (PnP) algorithm. There still difference distance from expected effect. To obtain more effective feature representation result, edge enhancement proposed to increase shape information object analyzing influence inaccurate matching on regression comparing effectiveness intermediate representation. Furthermore, although transformation matrix composed rotation translation matrices 3D model points 2D pixel points, two variables essentially different same network cannot be used both process. Therefore, improve PnP algorithm, this paper designs dual-branch predict information. Finally, method verified public LM, LM-O YCB-Video datasets. The ADD(S) values 94.2 62.84 LM datasets, respectively. AUC ADD(-S) value 81.1. These experimental show that superior similar methods.