Abstract 
In this seminar we start by a looking at an interesting method, boolean map saliency detection, first proposed by Zhang et al 2013. The idea behind this method is simple: Iteratively threshold an image using randomly selected threshold values, and count how often each pixel is disconnected from the image border in the resulting binary segmentation (referred to by Zhang as a boolean map). The resulting perpixel probability was shown to give a very good predicition of eyefixation, e.g., parts in the image that an observer is likely to look at. It turns out that the output of this procedure can be computed exactly and efficiently, without performing any actual random thresholding segmentations. Recently, we have found that similar results can be obtained also for other, similar stochastic thresholding procedures. The key finding is that there is a close connection between stochastic boolean maps and a concept from mathematical morphology called component trees.
