Leveraging synthetic data from CAD models for training object detection models – a VR industry application case

Abstract

In this paper we evaluate the applicability of using synthetic data, based on computer aided design models, to automatically detect objects in the real world. The aim is to enable scalable deep learning-based object detection to track and identify physical objects using a single low-cost camera. The approach is demonstrated and evaluated through a case-study involving a physical scale-model of an industrial plant connected to a virtual environment, aimed at facilitating multidisciplinary collaboration and immersive visualization. The digital models are simulated using domain randomization, and subsequently used to train object detection models. The results show the methods’ ability to generalize to real data, with accuracies up to 87%, demonstrating the scalability of the approach. Potential applications in industry are discussed based on these results.