Formally capturing spatial semantics is a challenging and still largely unsolved research endeavor. Qualitative spatial calculi such as RCC-8 and the 9-Intersection model have been employed to capture humans' commonsense understanding of spatial relations, for instance, in information retrieval approaches. The bridge between commonsense and formal semantics of spatial relations is established using similarities which are, on a qualitative level, typically formalized using the notion of conceptual neighborhoods. While behavioral studies have been carried out on relations between two entities, both static and dynamic, similar experimental work on complex scenes involving three or more entities is still missing. We address this gap by reporting on three experiments on the category construction of spatial scenes involving three entities in three different semantic domains. To reveal the conceptualization of complex spatial scenes, we developed a number of analysis methods. Our results show clearly that (I) categorization of relations in static scenarios is less dependent on domain semantics than in dynamically changing scenarios, that (II) RCC-5 is preferred over RCC-8, and (III) that the complexity of a scene is broken down by selecting a main reference entity.