This is an imperfect but interesting paper. The reviewers discussed it following the rebuttal, and while some of their concerns were addressed, it was agreed that the paper would be made stronger with a more thorough evaluation. Additionally, the dataset collection/training description is not clear, and it was felt this part of the paper would benefit from a rewrite. That said, the approach is novel and interesting, and the argument could be made that it is better to publish it now, leaving further applications for future work, than to expect perfection from every conference publication. As such, I am happy to recommend acceptance, but the authors should ensure they highlight the experimental (and other) limitations of their approach and its evaluation clearly in the final version, and make significant improvements to the writing (especially regarding the dataset) based on the feedback of the reviewers. Key strengths of the paper: * original method and application * interesting results * will prompt discussion and follow up work