The Recognition and Representation of 3D Images for a Natural Language Driven Scene Analyzer

Two necessary components of any image understanding system are an object recognizer and a symbolic scene representation. The LandScan system currently being designed is a query driven scene analyzer in which the user's natural language queries will focus the analysis to pertinent regions of the scene. This is different than many image underderstanding systems which present a symbolic description of the entire scene regardless of what portions of that picture are actually of interest. In order to facilitate such a focusing strategy, the high level analysis which includes reasoning and recognition must proceed using a topdown flow of control, and the representation must reflect the current sector of interest. This thesis proposes the design for 3 goal-oriented object recognizer and a dynamic scene representation for Landscan - a system to analyze aerial photographs of urban scenes. The recognizer is an ATN in which the grammar describes sequences of primitives which define objects and the interpreter generates these sets of primitives. The scene model is dynamically built as objects are recognized. The scene model represents both the objects in the image and primitive spatial relations between these objects.