An Appearance-based Approach to Object Recognition in Aerial Images

This report explores the design of \appearance-based" object recognition systems and their application to the analysis of aerial imagery. Our system, in contrast with the more prevalent model-based vision systems, in which explicit three dimensional models of objects and physics-based reasoning about the image formation process are combined, models objects in terms of how they appear to a computer vision system (i.e., what their structure is in possible segmentations of images) and identiies objects using combinatorial search constrained by the system's knowledge of and experience with nding the objects in previous supervised and unsupervised situations. As in other recognition systems, the \appearance-based" object recognition system that we have designed includes a low-level vision element, in which the image is preprocessed and segmented into components. The segmentation produced is a hierarchical one, with the hierarchy based on border contrast. The high-level vision phase uses the hierarchy and the \appearance-based" models of objects to heuris-tically combine components from various levels of the hierarchy into possible instances of objects. These are further analyzed by shape delineation processes and a nal veriication selects the combinations that correspond to locally best instances of objects.