Evaluation of computational attention operators using human image recognition
Sandra. Polifroni · eScholarship@McGill (McGill) · 2000
This thesis presents a novel method of evaluating computational attention operators, which select locations of interest in an image, using a human image recognition task. Assuming that locations which are maximally interesting will be most useful for recognizing an image, it follows that a location selected by an attention operator will facilitate image recognition if it is of interest to a human. Since attention operators are increasingly being used to replace humans in vision tasks, it is relevant that their performance be compared to human vision. Five different operators were evaluated. Human subjects were shown a series of black and white images in quick succession after which they were presented subimages extracted from the original image set as well as from other images. Subjects were asked to indicate whether they could recognize the subimages. The number of false-positives and true-positives associated with each operator provided information on the interest of the selected locations. Results show that the operators do not perform equally, with some selecting more recognizable image locations than others.