Synthetic data generation is an appealing approach to generate novel traffic scenarios in autonomous driving. However, deep learning techniques trained solely on synthetic data encounter dramatic performance drops when they are tested on real data. Such performance drop is commonly attributed to the domain gap between real and synthetic data. Domain adaptation methods have been applied to mitigate the aforementioned domain gap. These methods achieve visually appealing results, but the translated samples usually introduce semantic inconsistencies. In this work, we propose a new, unsupervised, end-to-end domain adaptation network architecture that enables semantically consistent domain adaptation between synthetic and real data. We evaluate our architecture on the downstream task of semantic segmentation and show that our method achieves superior performance compared to the stateof-the-art methods.
I. INTRODUCTIONAutonomous vehicles employ machine learning techniques in order to understand surrounding environments. This requires high generalization performance of the perception subsystem of an autonomous vehicle towards the environments and traffic scenarios it might encounter in the real world. That includes different variations of such environments (incl. lighting and weather conditions etc.), near-accident scenarios, and so-called long tail of events. Collecting sufficient training data that can cover various kinds of traffic scenarios in real-world environments is often not feasible. Moreover, such training data acquisition implies typically manual annotation. Such annotation might be a very laborious and timeconsuming task [3] especially in the use-cases like semantic segmentation as they require per-pixel labeling.Synthetic data generation is a promising approach to overcome the described problem of training data acquisition. It is a cost-effective method where the annotation of the data can be generated practically at no cost and variance of potentially simulated traffic scene scenarios has no limit. However, the semantic predictor trained with synthetic data reveals a catastrophic performance accuracy drop when evaluated is done on real data [23]. The rendering process introduces the artifacts known as the domain gap.Commonly, the gap problem is tackled by means of domain adaptation. Domain adaptation methods use generative networks based on adversarial training to translate
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.