Name of Reviewer ------------------ Peter Henry Key Contribution ------------------ This paper drives a closed form solution to extract the alpha channel from an image. By eliminating F and B, using relatively weak assumptions, the matte can be extracted with only scribbles from the user. Furthermore, the same cost function can be used to provide the user with hints to guide the placement of useful scribbles. Novelty -------- This seems to me to be novel work, particularly the derivation of a cost function only in alpha that can be minimized with relatively simple math. Also exciting is the use of eigenvectors (in a way I didn't really understand) to provide useful hints to the user as to the placement of foreground and background scribbles. Reference to prior work ----------------------- I am unaware of other work that should be cited. Clarity ------- Motivation, relationship to previous papers, and details and proofs are all present in well organized detail. All definitions and citations seem to be in order. I was somewhat intimidated by all the matrix math, but I suspect that someone with more vision experience would be more comfortable. Technical Correctness --------------------- I had difficulty following the proofs of the two theorems in the paper, and feel that I (personally) would need some examples (beyond what there was space for in the paper) to really understand the math. Experimental Validation ----------------------- I was very impressed with the experiments, particularly the quantitative comparison of the algorithms where 2000 subimages from a background were composited with foreground smoke, and the extracted matte was compared against ground truth. This seemed a fair and useful way of measuring performance. Several inspiring examples of the algorithm's performance were given, and shortcomings and assumptions were clearly stated. Overall Evaluation ------------------ I was very impressed with this paper's combination of novel method, rigorous proof, comparison to previous methods, and experimental technique. I would like to see the system work in person! Questions and Issues for Discussion ----------------------------------- I believe that a good direction to take this work would be combining the algorithm and "hints" together into an easy to use UI, with constant feedback provided by the system. It seems that the memory limitations of Matlab's solver were an unnecessary limitation to using this algorithm on larger images, and a custom solution should probably be explored. I would have liked to see some more examples of the algorithm making mistakes, so that the work could be extended (for example, where the assumptions about foreground and background "color linearity" did not hold).