Selecting re-used 3D model for object recognition by using handy camera

Abstract

To automate pick and place task by robots in warehouses is processed because of labor shortage. In order to pick and place items, the robots need to recognize the items. Some researches propose deep learning method for object recognition. They are trained by using simulated data. However, they require many 3D models of the items. Making 3D models in warehouses spends time and needs special equipment. Thus, to reduce the making cost is important for utilizing the deep learning method in real warehouses. In this study, we developed a method that re-uses primitive 3D model for building a 3D model of a new item. The method estimates a primitive class of the new item. Then the method builds 3D model corresponding the primitive class. We confirmed that the proposed method can make a 3D model of a box item and a cylinder item respectively.