Saliency and Localization by Nested Window Search

We model salient objects in images as being surrounded by more edges than they have in their interior, and address their detection by an algorithm that searches for optimal nested pairs of rectangular windows. The window that maximizes the sum of edge scores within the picture-frame like region between the two rectangles is returned. The space of all pairs of nested rectangles is a subset of R8, too large for exhaustive search even with branch and bound. We therefore restrict the geometry of the nested window pair and give an efficient algorithm to do the search in O(n1.5) running time for an image of n pixels. We demonstrate through experiments that nested window search is useful and fast in localizing objects of interest in a completely unsupervised way. Supervised classification may benefit from nested window search as well.