6D Pose Estimation of Transparent Objects Using Synthetic Data

Abstract

Transparent objects are one of the most common objects in everyday life. The poses of these objects must be estimated to pick and manipulate such objects. However, recognizing and estimating the poses of transparent objects is still a challenging task in robot vision, even after the emergence of 3D sensors. It is difficult to detect transparent objects because of the absorption and refraction of light, and the appearance of transparent objects can vary in different backgrounds. In this work, we propose a simple yet effective method for transparent object pose estimation, in which we address the problem using a synthetic dataset to train a deep neural network and estimate the pose of known transparent objects. After creating a synthetic dataset for transparent objects, we used a one-shot deep neural network to estimate the 6D pose of a known object. To the best of our knowledge, this is the first time synthetic data have been used for transparent object pose estimation. We conducted experiments on 3D printed transparent objects in a real environment as well as in a simulation environment. The results show that the proposed method can successfully estimate the pose of a transparent object even though it is only trained using synthetic data.