Environmental Relations in Image Understanding: The Force of Gravity

Kender, John R.

This paper shows how assumptions and information concerning the external world properties of horizontal and vertical can aid in the analysis of images, even at the very lowest levels of processing. First, the author reviews the pervasiveness of the force of gravity, and its influence on most natural image understanding systems. Next, he derives several fundamental mathematical results relating phenomena in both the gradient space and the images space to the external world attributes of horizontal and vertical. He then shows how these results interrelate three imaging phenomena: the surfaces in the image, the external sensor parameters, and the environmental labels. It is detailed how, in general, specific information regarding any two of these phenomena can be used to quantitatively derive the third; occasionally one can do even better. Algorithms for such quantitative derivations are presented, including two based on the Hough transform. The author further shows how certain environmental perpendicularities can be exploited very efficiently, and even elegantly: ordinarily complex math simplifies to the extent that environmental distances can be directly read off the image. The power of such environmental labels is then demonstrated by an analysis of the source of ambiguity in a simple illusion-like image configuration. The paper concludes with an analysis of the class of heuristics that have been invoked throughout. They are seen to be instantiations of the shape-from-texture meta-heuristics that "near implies preferred" and "preferred implies simple."



More About This Work

Academic Units
Computer Science
Department of Computer Science, Columbia University
Columbia University Computer Science Technical Reports, CUCS-047-83
Published Here
October 20, 2011