Using Generic Geometric Models for Intelligent Shape Extraction

Object delineation that is based only on low-level segmentation or edge-finding algorithms is difficult because typical edge maps have either too few object edges or too many irrelevant edges, while object-containing regions are generally oversegmented or undersegmented. We correct these shortcomings by using model-based geometric constraints to produce delineations belonging to generic shape classes. Our work thus supplies an essential link between low-level and high-level image-understanding techniques. We show representative results achieved when our models for buildings, roads, and trees are applied to aerial images.