Steps towards Autonomous Driving: Deep Semantic Segmentation among vehicle viewpoints

Giving the machines the ability to see and understand the overall context of what they are looking is the ultimate goal of Computer Vision. Simultaneously, Deep Learning aims at schematize the human brain structure building models that will learn from huge amount of data. From the combination of this two Machine Learning fields comes out lots of algorithms that reaches state-of-the-art performances on many tasks, including the one of the autonomous driving. It is important, in fact, that self-driving cars understand the environment where they are operating and that’s why have been developed the so-called semantic segmentation and domain adaptation networks. Both of this mentioned methods required huge amount of data to be trained and here is inserted the work done on this master thesis; indeed, it deals with the development of a completely new dataset for semantic segmentation, IDDA, characterized by its greatness and variety, allowing multi-source and domain-shift problem training and testing. The thesis also studies and tries to solve this problem, discovering that, till now, the available domain adaptation techniques, fails to completely close this gap and behave in an unexpected way when dealing with different viewpoints.