Name of Reviewer ------------------ Ryan Kaminsky Key Contribution ------------------ The main contribution of this paper is the development of a system for interactive digital matting that reduces the original problem to a simpler problem that is a quadratic function in only alpha, eliminating the need to use the foreground and background colors of the pixel as part of the system of equations to solve. They are able to produce results comparable to the best matting techniques available using only a linear system that can be solved much faster. Novelty -------- The novel portion of this paper is how they take the previous work and general equation I=aF+(1-a)B, but assume that F and B are locally smooth. Additionally, they introduce the notion of using eigenvectors to analyze the image in terms of the matting and in helping guide the user input scribbles to achieve the best matte. Reference to prior work ----------------------- Nothing to add here. Clarity ------- The paper is reasonably clear overall. I think more motivation could be given to the problem, particularly for those who are unfamiliar with the subject. Also, more information on whey this is different from segmentation would be helpful. I think the technical portion could use some clarity, this is addressed in the section. I think the transition to section 4 could use some work as this is an interesting part, but seems to be not as well setup as other sections of the paper. Technical Correctness --------------------- The paper sometimes introduces new mathmatical equations but could go into more depth about how they are formed. The derivations are a bit difficult to follow because of this. Experimental Validation ----------------------- This paper shows examples of several experimental images as well as much data about all of the experiments. This method seemes to beat all other methods in all cases. I think they should have shown cases where the Wang-Cohen algorithm is superior to theirs. They do show a failure case for the algorithm, but the don't show how Wang-Cohen does in this case. Also, when comparing the action figure in figure 7, they show that adding a few additional scribbles significantly improves the matte, but they don't show the efect of these additional scribbles on the other algorithms. It is impressive that the results are so much improved, but the comparison isn't fair. They say that their results are competitive with more complicated, non-linear cost functions, but all the evidence they present shows that their algorithm is superior to all others. Overall Evaluation ------------------ This paper is interesting as it presents an algorithm for matting that is competitive with other leading algorithms, but is faster and involves linear equations instead of non-linear. The derivation of the equations could be explained a little more fully. Also, I think the experiments could have focused more on the differences between this algorithm and the other leading algorithm. Questions and Issues for Discussion ----------------------------------- How will this algorithm work on non-natural images? If images are very noisy or the assumption that they are smooth by the small windows doesn't hold, how will the algorithm work? Can eigenvectors be used to "automatically" apply the sketching that the user is supposed to apply. How sensitive is the algorithm to the user input? Would the average user be able to scribble correctly.