First, I'd like to say that I thought it was very interesting that by posing the segmentation problem as a Markov random walk problem they were able to formulate a framework where the could perform some sort of supervised learning. I would have liked to have seen a bit more on the effectiveness of this learning and what sort of segmentation "features" could be found.

Another thing that has become apparent to me looking at the last few papers is that it would be very nice to develop some sort of standardized test set for examining the performance of image segmentation techniques. Does anyone know of any such set? I partially understand why one doesn't exist given that segmentation may be a more subjective metric. Given an image, there may not be one right way to segment it. I think the papers are beginning to do a good job of theoretically showing why these methods work, but I think an experimental analysis would allow insight into how these methods might work in a practical application.