Name of Reviewer ------------------ Jacki Roberts A Closed Form Solution to Natural Image Matting Key Contribution ------------------ Building an alpha matte using only a few scribbles input by a user to determine which pixels in an image belong to the foreground and which belong to the background. This method is more efficient than most state-of-the-art complex methods, and more accurate than most quick methods. Further, it provides a closed-form solution using only the alpha (transparency) values, independent of the actual foreground and background pixel colors. Novelty -------- Producing efficient yet high-quality alpha mattes, while minimizing the user input required. Eliminates the actual fore- and background color data when solving for the alpha unknowns. Appears to be the first paper to provide a closed-form solution. Also appears to be the first method that provides Eigenvalue-based feedback to the user to refine the scribbles used to create the alpha matte. Reference to prior work ----------------------- Requires seemingly strong assumptions about foreground and background smoothness for the mathematical substitutions to be valid. Justifies this by claiming all prior methods make similar assumptions. Clarity ------- They authors were clear about their method, how it differs from related methods, and stated the assumptions their method relies on. They were also clear about how they constructed their test images. Technical Correctness --------------------- I didn't verify the precise details of their matrix derivations, but the theorems themselves were correct assuming these derivations were correct. The interpretation of the mathematical results (e.g. which grey-area pixel groups are associated with smaller Eigenvalues) seems correct as well. Experimental Validation ----------------------- This seemed the weakest part of the paper. Their claims related to figure 5, that their results are comparable to other methods, seemed to be based solely on visual examination and not on error data. Their figure 6 shows good results, but on a very artificially constructed set of images. Although there were 4000 composite images, they all involved a homogeneous foreground superimposed on backgrounds all derived from a single image (whose contents are distinct from the two foregrounds). It felt a bit as if the experiments were too strawman-like, constructed to produce results that made their method stand out from the others. Overall Evaluation ------------------ The paper was good but needed a little more depth. They weren't completely clear on how justified they were to make their assumptions, except to say that most natural images seemed to adhere to the color line model. Their experiments either used pretty specific images (white flower petals on a very distinct background) or artificial images superimposed on subsets of a single natural image. Hence, they didn't spend much time characterizing the edge cases where their method could fail, beyond showing one in figure 9. Their smoothness assumptions, like the color line model, are consistent with the assumptions made by state-of-the-art methods, but still seem rather strong. Questions and Issues for Discussion ----------------------------------- How could this method apply to more complex backgrounds and foreground? They drop color variables from the functions that build their alpha matte--could losing this potentially valuable data be a problem in more complex images? They stressed the efficiency of their algorithm, which implies it would be most useful in a real-time image or video analysis. What value does it provide over iterative approaches that refine F,B and alpha in non real-time applications?