Saliency detection via divergence analysis: a unified perspective

Computational modeling of visual attention has been a very active area over the past few decades. Numerous models and algorithms have been proposed to detect salient regions in images and videos. We present a unified view of various bottom-up saliency detection algorithms. As these methods were proposed from intuition and principles inspired from psychophysical studies of human vision, the theoretical relations among them are unclear. In this thesis, we provide such a bridge. The saliency is defined in terms of divergence between feature distributions estimated using samples from center and surround, respectively. We explicitly show that these seemingly different algorithms are in fact closely related and derive conditions under which the methods are equivalent. We also discuss some commonly-used center-surround selection strategies. Comparative experiments on two benchmark datasets are presented to provide further insights on relative advantages of these algorithms