Investigation of an Integrated Synthetic Dataset Generation Workflow for Computer Vision Applications

Object detection and other machine learning technology applications play an important role in various areas of computer vision (CV) applications within the product lifecycle, especially in quality assurance or general assembly assistance. While the implemented CV-based systems provide great benefits, training and implementing deep learning models is often a tedious and time-consuming task, especially in the field of object detection. To accomplish good results, large datasets with a high quantity of object instances in a bright variety of poses are required These are generally created manually and are therefore very time consuming to create. To improve the training process, synthetic training data can be used. It is generated within a virtual environment using a product’s geometry model. In this paper, the authors propose a synthetic dataset generator for object detection, that is integrated into a PLM system to automate the process of collecting and processing the CAD data for creating the synthetic machine learning training dataset. Domain randomization is used to eliminate the effort of creating a virtual environment, to fully automate the dataset generation, and to increase the generalization of the model. The trained detector is tested on an object detection demonstrator set-up to evaluate its performance in a real-world use case. For evaluation purposes, the authors also provide a comparison of the test results to an object detection model that is trained without domain randomization, using a very close-to-reality virtual environment.