Generating images with physics-based rendering for an industrial object detection task: Realism versus domain randomization

Abstract

Limited training data is one of the biggest challenges in the industrial application of deep learning. Generating synthetic training images is a promising solution in computer vision; however, minimizing the domain gap between synthetic and real-world images remains a problem. Therefore, based on a real-world application, we explored the generation of images with physics-based rendering for an industrial object detection task. Setting up the render engine’s environment requires a lot of choices and parameters. One fundamental question is whether to apply the concept of domain randomization or use domain knowledge to try and achieve photorealism. To answer this question, we compared different strategies for setting up lighting, background, object texture, additional foreground objects and bounding box computation in a data-centric approach. We compared the resulting average precision from generated images with different levels of realism and variability. In conclusion, we found that domain randomization is a viable strategy for the detection of industrial objects. However, domain knowledge can be used for object-related aspects to improve detection performance. Based on our results, we provide guidelines and an open-source tool for the generation of synthetic images for new industrial applications.