“…In the literature this problem has been mainly approached from the tag ranking perspective. In the generative methods, which involve topic models [3,42,60,44] and mixture models [32,25,53,15,6,13], the candidate tags are naturally ranked according to their probabilities conditioned on the test image. For the non-parametric nearest neighbor based methods [37,38,35,27,22,34,61], the tags for the test image are often ranked by the votes from some training images.…”