Tag Ranking is a paper from Microsoft which discusses how to rank the flat collections of tags associated to Flicker images. The key idea is to weight tags according to (a) their probabilistic relevance (a slightly sophisticated variation of well-known TF*IDF, where the probability density function is estimated with a KDE approach) and to (b) a random walk on a tag similarity graph.
The results obtained on a limited testbed are encorauging, but I want to see how this scales on a large database of tagged images. In addition, I would have leveraged the taggers and not just the tags... Maybe an idea for a future work paper.
No comments:
Post a Comment