This was exactly what we were trying to do with ReSPEC. Finding consistency in an unlabeled set of images, and extracting that consistency (object). I have no real comment other than to say this paper is awesome and I will respectfully let them use the name ReSPEC for their method, so long as we get cited :) LONG LIVE ReSPEC

Ignoring the fact that training takes so long, this seems like the perfect application to learning a map from keywords for an image search query to a segmentation tree that is representative of most of the objects in the set of images.

How about applying this approach to Grozi? We have a grocery list and want to learn a representation for these objects without doing much work. We can feed each grocery item into an image search engine, get the top N results, feed them into this framework, and get out a representation for the grocery item. Make sense?