Learning Qualitative Spatial Relations for Object Classification

In this paper, we describe research into learning a model of the general 3D spatial relationships between objects found in human environments. Our approach trains a maximum entropy model using the qualitative spatial relationships between all the objects in several environments and produces a model of the underlying spatial regularities between the objects. It is tested as a purely spatial object classifier on the task of recognizing hundreds of objects based on their position relative to the other objects in their environment. We also introduce a novel source of data about object arrangement in human environments from the commercially released computer game Elder Scrolls 4: Oblivion.