Comparing Saliency Maps with Eye Movement Data

Saliency maps are used to identify the most visually attractive portions of an image [1]. They have been successfully employed in computer graphics to segment images into regions the user is most likely to focus upon. These regions are then selectively rendered in higher quality than less salient portions, at a reduced overall computational cost, but without the viewer being aware of this quality difference [2].

We wished to ascertain if it was feasible to use these same techniques to detect salient areas on Web pages, and to then test whether the areas identified by the saliency algorithm match the areas of the page on which people are most likely to fixate.

The images below show the standard page, a saliency map of the page, and the saliency map superimposed on gaze hotspot data gathered during an eye tracking study. The map shows that bold and colourful areas are predicted to be the most salient. Although the match is not exact, a link can be seen between salient areas, and the areas that received the most fixations. We plan to investigate this relationship further.

[1] L. Itti, C. Koch, and E. Niebur. Model of saliency-based visual attention for rapid scene analysis. Pattern Analysis and Machine Intelligence, 20:1254–1259, 1998.

[2] V. Sundstedt, K. Debattista, P. Longhurst, A. G. Chalmers, and T. Troscianko. Visual attention for efficient high-fidelity graphics. In Spring Conference on Computer Graphics, May 2005.

Directed attention home

Tracking eye movements on the BBC News Website

Using multimodal cues to direct the user's attention

The directed attention project is funded by the EPSRC under grant EP/D036518/1 (Sept 2005 - July 2006).