Shape Enhanced Keypoints Learning with Geometric Prior for 6D Object Pose Tracking

Until now, there has not been much research in exploiting geometric reasoning on object shape and keypoints in object pose estimation. First, the current RGB image and quaternion representing rotation in the previous frame are fed to a multi-branch neural network responsible for regressing sparse object keypoints. The initial object pose is estimated using PnP, which is adjusted in a least-square optimization. The weights of boundary and keypoints components are determined in each iteration via geometric reasoning on the projected and segmented 3D object boundary, object shape extracted by a pretrained neural network and keypoints extracted by our network. Different from previous methods, our voting scheme is object boundary-based. We demonstrate experimentally that the accuracy of pose estimation is competitive in comparison to the accuracy of SOTA algorithms achieved on challenging YCB-Video dataset.