Abstract

Machine-vision research shows that accurate three-dimensional (3-D) reconstruction of objects is difficult. Human-vision research shows that human judgments about 3-D features (depth and volume) are not accurate. This converging evidence, along with the known human ability to recognize shapes, suggests that accurate recognition of 3-D shapes is possible without accurate reconstruction. We consider only a part of this general problem, namely, recognition of 2-D shape in 3-D space. The proposed method is based on a decomposition of the differences between a presented and a standard shape into two components. One component is the amount of perspective difference (i.e., the difference produced by perspective transformation), and the other is the amount of nonperspective difference. An estimate of tilt is necessary for this decomposition, although the sensitivity of recognition to errors in tilt is small. The magnitude of the nonperspective difference is a measure of dissimilarity between shapes (if the dissimilarity is small, the shapes are considered to be identical), and the magnitude of perspective difference is an estimate of slant. Comparison of simulation results with psychophysical results suggests that the method is a plausible model of human shape recognition.

© 1990 Optical Society of America