Integrating habituation into saliency maps

Computational modelling of visual saliency has been an active area of research since the last two decades. It has profound impact on robotics as it helps in automatically selecting visually interesting regions from a given scene. A model of visual saliency generates a saliency map - which is a two-dimensional representation of the visual scene - where each pixel highlights the degree of saliency at a given spatial location. A saliency map is meant to be a surrogate representation of the spread of visual attention over the scene. Several saliency maps such as [1], [2] and [3] have been proposed in the literature. They are successful in mimicking the human eye-gaze on images for a free-viewing condition. The saliency models operate at the pixel level and rely on colour contrasts to compute the saliency at a given point. However, humans do not necessarily view images at the pixel level, but rather segment the scene in terms of objects. This is at a higher level of abstraction which the contemporary saliency models do not handle.