Computing Human-Like Qualitative Topological Relations via Visual Routines

A core problem in spatial reasoning is finding an appropriate set of relationships to compute. This paper proposes that humans represent topological relationships between 2D regions using three basic, qualitative relations: contains, intersects, and overlaps-with. We show how these relations can be computed from sketched inputs using a model of mid-level perception. Results from a pilot experiment indicate that these three relationships suffice to explain people‟s judgments on four English spatial terms (“intersects”, “overlaps”, “connects to”, and “contains”), although a combination of the three is generally required for each term.