A great book came out,

Nearest-Neighbor Methods in Learning and Vision:
Theory and Practice

and Chapter 6 talks more about this!!

http://people.csail.mit.edu/gregory/annbook/book.html

Could you clarify how one chooses the set C of candidate objects? Performance probably depends on the selection of such candidates.
Thanks

Just out of curiousity, as long as we're giving coorespondences for training. Can anyone convince me that this sort of combination of embeddings would outperform something simple like LFD where ach class is a correspondence and each image is a sample in all classes?