Visual Search

Predicting eye fixations is typically done with some form of the saliency map, e.g. (Itti, Tatler, Rajashekar). However the structural analysis of those approaches is rather sparse and can not explain why humans for instance foveate near intersecting lines (see below). In contrast, the structural decomposition I have developed, can for instance explain all the pop-out phenomena observed in human visual search, by simply taking the variance of the vector descriptors (contour & areas). To apply the methodology to fixation prediction as the one below, I need to elaborate it a bit more, by for instance creating grouping algorithms. Together with Ben Tatler at the University of Dundee, I have also started to analyse saccadic target selection in natural scenes.

If shown a line drawing, humans tend to foveate clusters of intersecting lines (from Noton & Stark 1971):