Marius Leordeanu
POLITEHNICA Bucharest and Simion Stoilow Institute of Mathematics of the Romanian Academy

The world can be interpreted from many perspectives and levels of semantic understanding, and all these interpretations are highly interdependent. We can often tell what happens in one part of a scene, at a specific time and location, by knowing what takes place in other parts. One way of interpreting the world can tell us a lot about other ways: geometry, semantics, motion and measurements from different sensors are strongly interrelated, each providing contextual information for the others.

In my talk, I will present our work during the last five years towards developing a complex multi-layer system that can exploit the connectivity between the many layers of sensing and understanding the world. Our multi-layer hypergraph is able to learn by itself, deal with missing data, estimate its own uncertainty and provide robust estimations. It achieves all these by reaching consensus among its many different layers through an iterative estimation and learning process.