Delineating a prominent object in an image by computer is an important step toward fully automated image analysis. This segmentation problem has been elusive and no viable solution has been found. A recent trend is to provide multiple images that contain the same class of objects of interest and design an algorithm to delineate objects that share common characteristics across images. This so-called co-segmentation problem is the topic of this paper.

The proposed method first derives a rough initial segmentation using an existing segmentation method and some heuristics. The initial segmentation is used to derive foreground and background statistical models across the images; the foreground shares the common characteristics across images while the background does not. It then applies an efficient graph search on each image to extract a region that agrees best with estimated foreground and background models. The segmented region is then used to refine the foreground and background models. The procedure repeats until convergence.

The paper provides an extensive set of empirical results and comparisons with other methods. The results look impressive. However, it provides neither implementation details nor examples of the initialization scheme. Since their iterative search is highly sensitive to the initial state, crucial information is missing. Also, there are many ambiguous notations, missing definitions, and unsubstantiated claims, which make the paper difficult to follow. More careful edits would have made the paper more readable and logically sound. The figures are well crafted and helpful. Overall, the paper is interesting, but leaves many questions.