On an elementary definition of visual saliency

Various approaches to computational modelling of bottom-up visual attention have been proposed in the past two decades. As part of this trend, researchers have studied ways to characterize the saliency map underlying many of these models. In more recent years, several definitions based on probabilistic and information or decision theoretic considerations have been proposed. These provide experimentally successful, appealing, low-level, operational, and elementary definitions of visual saliency (see eg, Bruce, 2005 Neurocomputing 65 125 - 133). Here, I demonstrate that, in fact, all these characterizations provide essentially the same measure of saliency. Moreover, where the original formulations rely on empirical estimates of the underlying probability density of low-level pre-attentive features, I show that saliency can be expressed as a closed-form solution based on purely local measurements and, surprisingly, without the need to refer back to previously observed data. Furthermore, it follows that it is actually not the statistics of the visual scene that would determine what is salient but the low-level features that probe the scene.