Generating Datasets from 3D CAD Models for Object Detection

Abstract

It is time-consuming to prepare training and testing data for object detection by using deep learning. In this paper, we proposed a method to generate the training data using CAD models instead of using real objects. To achieve this, we needed to convert the CAD models into point clouds. Then, the objects to be detected also needed to be converted into the point-cloud format. The key to obtaining the point clouds of the objects was to find their masks using depth images captured by a depth camera. After the mask of all the objects were available, we separated the mask into each object's mask. The separated mask was then used on the depth image to obtain the object's point cloud for object detection. Using the proposed method, it was possible to use 3D CAD data to quickly train a deep learning model to detect objects. Our preliminary results showed that the accuracy of object detection reached about 89%.