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 biologicallyinspired models, followed by the information-theoretic and recently statistical-based models. This dissertation looks at a state-of-the-art statistical model and studies what effects the histogram construction method and histogram distance measures have on detecting saliency. Equi-width histograms, which have constant bin size, equi-depth histograms, which have constant density per bin, and diagonal histograms, whose bin widths are determined from constant diagonal portions of the empirical cumulative distribution function (ecdf), are used to calculate saliency scores on a publicly available dataset. Crossbin distances are introduced and compared with the currently employed bin-to-bin distances by calculating saliency scores on the same dataset. An exhaustive experiment with combinations of all histogram construction methods and histogram distance measures is performed. It was discovered that using the equi-depth histogram is able to improve various saliency metrics. It is also shown that employing cross-bin histogram distances improves the contrast of the resulting saliency maps, making them more perceptually meaningful but lowering their saliency scores in the process. A novel improvement is made to the model which removes the implicit center bias, which also generates more perceptually meaningful saliency maps but lowers saliency scores. A new scoring method is proposed which aims to deal with the perceptual and scoring disparities