On the knowledge required to label a picture graph

Abstract

I analyze the information content of scene labels and provide a measure for the complexity of line drawings. The Huffman-Clowes label set is found to contain surprisingly little additional information as compared to more basic label sets. The complexity of a line drawing is measured in terms of the amount of local labeling required to determine global labeling. A bound is obtained on the number of lines which must be labeled before a full labeling of a line drawing is uniquely determined. Methods are provided for obtaining subsets of lines whose labeling is sufficient to imply the labeling of the remaining lines. I present an algorithm which combines local sensory probing with knowledge of labeling constraints to proceed directly to a labeling analysis of a given scene.