Synth2Real – Training Neural Networks with Virtual Data

Prof. Dr. Angela Dai, Technische Universität München

In recent years, we have seen stunning progress in understanding visual data using modern machine learning techniques. However, much of this success in visual understanding is based on supervised machine learning and thus relies on the availability of large, annotated training data sets. Using synthetic data for training provides a unique opportunity for both generating data at scale and for unusual situations. This project addresses visual understanding tasks by developing new hybrid learning methods to enable transfer learning between synthetic and real data domains, thereby changing the way we currently train neural networks.