Break and conquer: Efficient correlation clustering for image segmentation

We present a probabilistic model for image segmentation and an efficient approach to find the best segmentation. The image is first grouped into superpixels and a local information is extracted for each pair of spatially adjacent superpixels. The global optimization problem is then cast as correlation clustering which is known to be NP hard. This study demonstrates that in many cases, finding the exact global solution is still feasible by exploiting the characteristics of the image segmentation problem that make it possible to break the problem into subproblems. Each sub-problem corresponds to an automatically detected image part. We demonstrate a reduced computational complexity with comparable results to state-of-the-art on the BSDS-500 and the Weizmann Two-Objects datasets.