Mask-based Object Pose Estimation with Domain Transfer

Object pose estimation is important for robots to understand and interact with the real world. This problem is challenging because the various objects, clutter and occlusions between objects in the scene. Deep learning methods show better performances than traditional problems in this problem but training a convolutional neural network needs lots of annotated data which is expensive to obtain. This paper proposes a general method by using domain transfer technology to efficiently solve object pose estimation problem. Besides, the proposed method obtains mask to achieve high quality performance by combing an instance segmentation framework, Mask R-CNN. We present the results of our experiments with the LineMOD dataset. We also deploy our method to robotic grasp object based on the estimated pose.