SpaRel - Using Spatial Relations for parts-based scene recognition and prediction


The goal of the proposed project is to utilize a system’s knowledge about spatial relations of objects or object parts to guide the visual scanning of the current scene. Here, characteristic patterns can be used e.g. to speed-up visual search, to recognize objects or to determine the context of the given scene.

The main focus of this project is on the recognition and interpretation of a visual scene which will benefit a lot by incorporating spatial relations of scene parts. For example, when seeing a car right of the street, it can be assumed to be parked, whereas the same car on the street can be assumed to be driving. Furthermore, the process of scanning a visual scene can also be improved by constraining the spatial search space. This means, that when detecting a fork in the scene a table can be assumed beneath as well as dishes or a knife might be seen right of this fork. A suited representation has to be found that allows the storing and learning of these spatial relations between object-parts or objects themselves.