One characteristic that sets humans apart from modern learning-based computer vision algorithms is the ability to acquire knowledge about the world and use that knowledge to reason about the visual world. Humans can learn about the characteristics of objects and the relationships that occur between them to learn a large variety of visual concepts, often with few examples.
The use of structured prior knowledge in the form of knowledge graphs can be very useful in this sense. Exploiting this kind of knowledge can foster performances on image classification task, as well as in different task.
For instance, semantic/functional priors incorporated into a deep reinforcement Learning framework can help to efficiently search and navigate in novel scenes. This approach can be applied in real world scenarios or in simulted ones, for instance in Interactive 3D Environments for Visual AI. Through this seminar we try also to give a brief overview of available 3d ambient simulators, hopefully usefull for the SAILab Vision Project.