Automatic Generation Method for Training Dataset of Instance Segmentation for Irregularly Shaped Object Picking

Abstract

In the field of robotics, robotic bin picking has been studied extensively. In this paper, we consider an irregularly shaped object as a grasping target. For detecting irregularly shaped objects, instance segmentation based on deep learning is one of the promising methods. One of the challenges for deep learning-based methods is how to reduce the time and effort to prepare the dataset for training. In this paper, we propose a method to automatically generate a dataset for learning instance segmentation using only information available from public image databases. The proposed method achieves mean average precision (mAP) of 0.85 for the automatically generated test data. It also showed mAP of 0.65 for the test data generated using untrained irregularly shaped objects, and achieved a success rate of more than 98% in picking experiments with the robot.