Analysis of Eye Fixations Data

Ali Alsam and Puneet Sharma


Eye fixations, Saliency


In this paper, we analyzed eye fixations data obtained from 15 observers and 1003 images. When studying the correlation matrix constructed based on the fixations data of one observer viewing all images, it was observed that 23 percent of the data can be accounted for by one eigenvector. This finding implies a repeated viewing pattern that is independent of image content. The examination of this pattern revealed that it was highly correlated with the center region of the image. The presence of a repeated viewing pattern raised the following question: Is our visual attention driven by salient image content? In order to answer this question, we analyzed the data across different observers viewing the same image. Our analysis showed that there was good agreement among observers for images containing people, faces, and text while poor agreement was observed for complex images such as landscapes, buildings, and street views. Our findings suggest that strong agreement between observers was due to top-down features of the image, i.e., context driven rather than bottom up features associated with low level image attributes.

Important Links:

Go Back