One aspect of the method which deserves further attention is that of modelling visual attention. While our formal definition proved successful for our particular application, it was created with implementation in mind; for example, the convolution window used for measuring edge density is a circular step operator, which allows for faster convolution, but has poor frequency-domain properties in comparison to a Gaussian operator. As such, it would be valuable to explore the behaviour of the method given a Gaussian convolution operator, and it would be further edifying to study the scale-space properties of the operator in general. More generally, we have only considered edge density for our model of attention. In keeping with the theme of distinctiveness, it would be worthwhile to consider other measures of uniqueness, such as symmetry, or edge orientation. Indeed, some of these issues have been considered by Bourque and Dudek [12].