Augmentation of Explicit Spatial Configurations by Knowledge-Based Inference on Geometric Fields

A spatial configuration of a rudimentary, static, realworld scene with known objects (animals) and properties (positions and orientations) contains a wealth of syntactic and semantic spatial information that can contribute to a computational understanding far beyond what its quantitative details alone convey. This work presents an approach that (1) quantitatively represents what a configuration explicitly states, (2) integrates this information with implicit, commonsense background knowledge of its objects and properties, (3) infers additional, contextually appropriate, commonsense spatial information from and about their interrelationships, and (4) augments the original representation with this combined information. A semantic network represents explicit, quantitative information in a configuration. An inheritance-based knowledge base of relevant concepts supplies implicit, qualitative background knowledge to support semantic interpretation. Together, these structures provide a simple, nondeductive, constraint-based, geometric logical formalism to infer substantial implicit knowledge for intrinsic and deictic frames of spatial reference.