Full-range affinities for graph-based segmentation

Graph-based segmentation has become a major trend in image segmentation. A key issue in graph-based segmentation is how to build the affinity matrix. Among the previous methods, many successful ones only compute the pairwise affinities between adjacent pixels and superpixels without considering the nonadjacent ones. Thus, they often obtain unsatisfactory results when foreground is cut into several nonadjacent parts by background or shadows. In this paper, we propose a full-range affinities learning method for graph-based segmentation. Our method computes the affinities both between adjacent pixels and nonadjacent pixels, which are inversely proportional to the shortest connectivity paths. The experimental results demonstrate the superiority of the proposed approach comparing with existing popular methods.