Enforcing Point-wise Priors on Binary Segmentation

Non-negative point-wise priors such as saliency map, defocus field, foreground mask, object location window, and user given seeds, appear in many fundamental computer vision problems. These priors come in the form of confidence or probability values, and they are often incomplete, irregular, and noisy, which eventually makes the labelling task a challenge. Our goal is to extract image regions that are aligned on the object boundaries and also in accordance with the given point-wise priors. To this end, we define a graph Laplacian spectrum based cost function and embed it into a minimization framework. For a comprehensive understanding, we analyze five alternative formulations, and demonstrate that the robust function version produces consistently superior results.