This paper generated quite a bit of discussion among the reviewers and the area chair. There was widespread agreement that the paper improved the state of the art and yet was not free of shortcomings. The most prominent shortcoming which was mentioned during the discussion was the lack of evaluation of the quality of estimated kernels. While I agree with this concern, I do not see this as a reason to reject the paper and I believe the strengths of the paper outweigh this concern. I am therefore happy to recommend acceptance but ask the authors to address the comments of the reviewers in their final version. In particular, I ask the authors to address the issue of the quality of the blur kernels..