Efficient combination of probabilistic sampling approximations for robust image segmentation

Methods based on pairwise similarity relations have been successfully applied to unsupervised image segmentation problems. One major drawback of such approaches is their computational demand which scales quadratically with the number of pixels. Adaptations to increase the efficiency have been presented, but the quality of the results obtained with those techniques tends to decrease. The contribution of this work is to address this tradeoff for a recent convex relaxation approach for image partitioning. We propose a combination of two techniques that results in a method which is both efficient and yields robust segmentations. The main idea is to use a probabilistic sampling method in a first step to obtain a fast segmentation of the image by approximating the solution of the convex relaxation. Repeating this process several times for different samplings, we obtain multiple different partitionings of the same image. In the second step we combine these segmentations by using a meta-clustering algorithm, which gives a robust final result that does not critically depend on the selected sample points.