Improving road-safety for autonomous vehicles, through better algorithms for perception of expected and unexpected objects

We aim to improve road-safety for autonomous vehicles, through better algorithms for perception of expected and unexpected objects by fundamental new theory, algorithms and implementation (code) for weakly supervised semantic segmentation. We aim to create new methods for weakly supervised segmentation that improve reliability and greatly reduce the prohibitive labelling cost, making segmentation of unexpected objects possible, by extending current approaches with an end-to-end training framework, incorporating uncertainty-driven reasoning and improved generative approaches.