Investigating Top-Down Influences in Image Segmentation

The process of segmenting an input image refers to the task of determining coherent image regions which are likely to correspond to the existing objects in the image. An image segmentation algorithm follows some basic grouping rules and usually attaches a label to each pixel, denoting which region the pixel belongs to. One such way to model segmentation as a labeling problem is to consider an image as a graph in which the nodes correspond to image pixels and edges reflect pairwise similarities between the pixels. Then the labeling process can be simply formulated as a graph partitioning problem where each resulting subgraph denote an image segment. However, it should be noted that this formulation is a purely bottom-up segmentation model. If we think of natural images, achieving a perfect segmentation is quite hard simply because several ambiguities might be present in the image so a purely bottom-up segmentation methods is doomed to fail. To alleviate this problem, it is very important to combine bottom-up cues with top-down information [1, 2, 3] to resolve the ambiguities.