Using the earth mover´s distance for perceptually meaningful visual saliency

Visual saliency is one of the mechanisms that guide our visual attention, or where we look. This topic has seen a lot of research in recent years, starting with biologically-inspired models, followed by the information-theoretic and recently statistical-based models. This paper looks at a state-of-the-art statistical model and studies the effects of using a cross-bin histogram distance instead of the usually employed bin-to-bin histogram distance on saliency detection. Specifically, the earth mover´s distance (EMD) is used to replace the Bhattacharyya coefficient, and it is shown that the EMD provides a more continuous distance function for the high-dimensional colour histograms. This feature also facilitates the removal of an implicit center bias from the model in question. Both these modifications lower a standard set of saliency scoring metrics, which is attributed to the noisy and haphazard nature of the eye-tracking data. An explicit selection task is identified as a possible avenue from which more robust and meaningful saliency scores can be developed.