Deep-Learning Pipeline for Object Pose Estimation from an RGB-D Image
You chan No, Young-Woo Kim, Daegun Kim, Hyeon-Gyu Han, Young-ki Song, Duksu Kim · Journal of Institute of Control Robotics and Systems · 2021
We propose a novel deep learning pipeline for estimating the 6D pose of an object from an RGB-D image. Our pipeline comprises two separate pipes: object detection and pose estimation pipes. By employing the You Only Look Once model, we were able to easily build the object detection pipe and achieve high accuracy for the task while decreasing the complexity of pose estimation for the next pipe. For the pose estimation pipe, we adopted the novel CNN (convolutional neural network) architecture, which can be easily configured and used in industrial sites. We generated datasets for five objects, including parts used in industrial sites, using CAD models and computer simulation in a virtual world. Our pose estimation model achieved up to 99.6% (90.9% on average) accuracy in terms of shape similarity (i.e., ) between the CAD models rotated by the ground truth and the inference result for the datasets. These results validate the usefulness of the proposed deep learning pipeline and the CNN architecture for 6D pose estimation.