http://www.techcrunch.com/2008/04/27/google-experiments-with-next-generation-image-search/
http://www.nytimes.com/2008/04/28/technology/28google.html
http://www.www2008.org/papers/fp506.html
In this paper, we cast the image-ranking problem into the task of identifying "authority" nodes on an inferred visual similarity graph and propose an algorithm to analyze the visual link structure that can be created among a group of images. Through an iterative procedure based on the PageRank computation, a numerical weight is assigned to each image; this measures its relative importance to the other images being considered. The incorporation of visual signals in this process differs from the majority of large-scale commercial-search engines in use today. Commercial search-engines often solely rely on the text clues of the pages in which images are embedded to rank images, and often entirely ignore the content of the images themselves as a ranking signal. To quantify the performance of our approach in a real-world system, we conducted a series of experiments based on the task of retrieving images for 2000 of the most popular products queries. Our experimental results show significant improvement, in terms of user satisfaction and relevancy, in comparison to the most recent Google Image Search results.
|
Neven Vision... they SHOULD do something great with that! I'm sure you're dreaming of that. |
> I wonder what motivates Google to publicize this information, as it could tip of their competition?
As you say, it's so computationally expensive that even the Googlecomp can't do it usefully. Besides, the idea isn't new... Riya says they tried that before, but gave up because it was too demanding to their computers, while working only on toy examples. |
Agree with Veky. If the Google* can't scrounge up the computing resources necessary, there's really no risk of anyone else doing it. But in the mean time they can attract future employees and also maybe spur academic research to improve on the idea (which they can buy later with their ginormous stock price if they haven't solved it themselves in the mean time).
Plus, lots of Google employees like writing papers and giving talks. If you don't let them do what they like they'll leave.
*We don't really have a word yet for their worldwide computer as a whole, do we? |
Veky: The idea is new (or at least the idea to implement a PageRank-like algorithm in this context) and it's not that computationally intensive. The paper suggested finding the first 1000 results using the current Google Image Search and applying this new algorithm to those images. Since the results can be cached, I'm sure Google can compute the results for the most popular 10000 queries or so. |
I don't think these results can be cached usefully, because they intricately depend on _other_ results. It's not that the image X has VisualRank V(X), it's that the image X, _considered as an element of the set of images S,_ has VisualRank V(X,S). Considered as an element of some other set, it could have a very different VisualRank.
Ok, maybe you could cache them according to individual _query_ (and we know most queries are repetitive), but what to do with those that are not? "Sorry, your query is too original [link](learn more)[/link]. You'll have to wait more than 0.1 second for results. You can also [link]view results ranked randomly[/link]." ;-) |
This PageRank (or VisualRank) is query-deapendent. I don't think it will work very well for queries that are specific because the number of results is small and they're heterogeneous. |
Somehow making something work only for the top-X queries seems very un-Google-like (it doesn't scale well, and Google also told people this 25%-of-all-queries-are-new-thing, even when it got partly corrected later)... but who knows. |
Microsoft managed to process images and find the same objects regardless of lighting, angle of sight, ... The demonstration of the idea can be seen at Photosynth (http://labs.live.com/photosynth/).
|
Photosynth requires extra software as in a plugin to run right? Also its available on windows only. Google Image search is available on any OS/Browser. |
Sure, you are right. Photosynth itself is actually just a demo. The interesting part was done by pre-processing – arranging the images (they used Flickr) based on the motives – it seems they are quite advance at image recognition. |