Ing. Vojtěch Franc, Ph.D.

Deep decision models are learned end-to-end from a large number of examples of input annotated with the required output. The most popular supervised learning methods are based on complete and precise annotations, which are difficult to collect in large numbers. On the other hand, collecting incomplete or noisy annotations is often much cheaper. The thesis will investigate methods to learn deep models from weak annotations. The thesis will focus on developing theoretically grounded methods and on their applications in computer vision.