Abstract

Perceptual organization is the ability to group (segment) automatically imagery into meaningful parts, which may then be matched against known objects (i.e., recognition) or combined into a coherent whole (i.e., learning new objects). This ability is critical to any general-purpose vision capability, and yet despite much research into use of grouping criteria such as continuity, nonaccidentalness, and goodness of form, there has been surprisingly little progress. We believe that a robust perceptual organization capability cannot be achieved by use of such low-level image-based grouping criteria but rather must rely on knowledge of generic formative processes that act to product coherence in the 3-D world. We describe a working system that employs generic knowledge about the shape of man-made structures (that they may be described as the union of rectangles) and about the shape of plants (that they may be described as the union of circles) to segment aerial imagery into meaningful parts, which are then grouped into houses, roads, and trees.

© 1987 Optical Society of America