Twin Space Based Monocular Image Object Pose All-in-One Labeling Method

The multi-object pose estimation problem is one of the fundamental challenges in the fields of robotics and intelligent transportation. However, the current research on 3D pose estimation of rigid objects focuses on a relatively small scale, which leads to a shortage of practical applications in this field. In this paper, we propose twin space based monocular image object pose all-in-one labeling method, and publish a pose labeling tool, called LabelImg3D. We construct a twin space equivalent to the reality space and a 3D model of the real rigid object. After that, we place the real space image (primary projection) in the twin space so that the image taken by the simulated camera in the twin space (secondary projection) can match with the primary projection. Lastly, by moving and rotating the 3D model in the twin space, the object in the secondary projection image and that in the primary projection image are aligned in the image-space so that the poses of the object can be obtained. In this paper, we open source a labeling tool LabelImg3D (https://github.com/CongliangLi/LabelImg3D). The experimental results demonstrate that our method can achieve a translation accuracy of more than 85% and a rotation accuracy of more than 90% for the same type of object with little dimensional variation. In addition, our method only uses a monocular camera, which greatly reduces the difficulty of estimating the object’s 3D positional data.