Task- and knowledge-driven scene representation

In this thesis a flexible system architecture is presented along with an attention control mechanism allowing for a task-dependent representation of visual scenes. Contrary to existing approaches, which measure all properties of an object, the proposed system only processes and stores information relevant for solving a given task. The system comprises a short- and long-term memory, a spatial saliency algorithm and multiple independent processing routines to extract visual properties of objects. Here, the proposed control mechanism decides which properties need to be extracted and which processing routines should be coupled in order to effectively solve the task. This decision is based on the knowledge stored in the long-term memory of the system. An experimental evaluation on a real-world scene shows that, while solving the given task, the computational load and the amount of data stored by the system are considerably reduced compared to state-of-the-art systems.