Deep Semantic Segmentation across environments for Autonomous Driving

In the autonomous driving context it has always been crucial to be able to teach a vehicle to precisely identify each entity in an image, in order to take appropriate decisions in various scenarios. Semantic segmentation aims at doing so by classifying each individual pixel, but, as well as all approaches based on Deep Learning, it requires lots of data to effectively train a network. Collecting big amounts of labeled data is far from trivial, so synthetic and real datasets were created to overcome such scarcity. This is not enough though, as real datasets are still too small and not various enough to satisfy this need, whereas readily available synthetic datasets do not offer a wide variety of scenarios to choose from. Moreover, semantic segmentation approaches have troubles dealing with different domains, being unable to generalize well to a given unseen domain. For example, training with synthetic samples and testing with real ones leads to a drop in performance. In this thesis work, a new and very large synthetic dataset is proposed in order to provide the research community with a novel challenge where a wide choice of domains is available, such as different towns, atmospheric conditions and vehicle viewpoints. The dataset is tested with a special focus to the domain shift issue, proving how the current state-of-the-art networks struggle with closing the domain gap, and demonstrating that further research in this area is needed.