A solution to the feature correspondence problem inspired by visual scanpaths

The organization of features into ordered sets to allow comparison of visua l images is called the correspondence problem. When presented with a famili ar image, humans scan an image in a scene with a specific and distinctive p attern. We suggest that the visual system uses scanpaths to solve the featu re correspondence problem. A computer algorithm that simulates this process was developed and tested on different classes of images. In agreement with psychophysical data, the algorithm processed familiar images with idiosync ratic "eye movement" patterns and with fewer "eye movements" than unfamilia r patterns; additional fixation time was spent on unexpected patterns. This solution of the correspondence problem is well suited to computer vision s ystems. (C) 1999 Elsevier Science B.V. All rights reserved.