Thanks for describing MS-SIM - I've always been concerned about using L2 or similar metrics for image similarity. Is there any work on learning an image similar

Thanks for describing MS-SIM - I've always been concerned about using L2 or similar metrics for image similarity. Is there any work on learning an image similarity metric? I would think if you Mechanical Turk'd a bunch of A/B images (with or without perturbations of various sorts or even totally different images) and asked the user to score the similarity you should be able to learn a pretty interesting scoring.