Realism Adversarial Networks: Train Deep Networks with Simulated Data

P. Torrione, M. Hibbard
CoVar Applied Technologies, United States

Keywords: Deep learning, simulations, GAN

Simulated data can be used to generate new classes, environments, and scenarios much more rapidly than real-world data collections. However, complex machine learning architectures can easily discriminate real and simulated data, so training deep networks with mixtures of real and simulated data rarely improves network performance on real data. “Realism Adversarial Networks” (RAN) are a type of network that enable us to modify data generated by simulators to “REALize” it so that the data is indistinguishable from real-world data and improve performance. CoVar has previously demonstrated the effectiveness of RANs for DoD relevant scenarios in IR, camera imagery, and videos.