Abstract
We present a paradigm for discovering the outlines of arbitrarily complex cultural objects in aerial imagery. The approach starts with a low-level image partition and generic (as opposed to specific or template-like) object descriptions. We then use geometric reasoning and context knowledge to suggest corrections to the discrepancies between the segmentation boundaries and the object models. Finally, when the corrections appear consistent with the generic cultural object model, we re-segment the partition to produce new labeled regions with clear semantic interpretations. The general features of our approach appear to be applicable to a number of other domains.
