Computation of Interestingness maps.

<p>(<b>A</b>) 3 different stimuli are shown together with their actual fixation and interestingness maps (<i>second</i> and <i>third</i> rows). The first row depicts three stimuli belonging to Natural and Urban categories as they were shown during the experiment. <i>Second</i> and <i>third</i> rows depict the empirical saliency and interestingness maps overlaid on the gray scale version of the stimulus. Empirical saliency maps are probability maps that show the probability of a given location to be fixated. Similarly interestingness maps represent for a given location the probability of receiving an interestingness rating. These were obtained with the help of a pointer device by an independent set of human subjects (n = 35). These maps were treated the same way as low-level feature maps in order to compute D_KL values. (<b>B</b>) Joint distribution of interestingness values and a low-level feature (Cornerness) that was most strongly correlated with fixation locations (<i>Left</i> panel). This distribution shows the co-occurrence of a low-level image based feature with interestingness ratings of human subjects at all fixated and non-fixated locations. <i>Middle</i> panel represents the distribution of same variables at exclusively fixated locations. The posterior distribution represents the two-dimensional saliency function (<i>right</i> panel). Notice that the saliency increases nearly completely as a function of interestingness ratings (contour lines) and only marginally as a function of low-level feature values. The saliency is therefore mainly modulated by the interestingness value of a location.