Google VisualRank

Google VisualRank
Google researchers at a web conference in Beijing announced they work on some kind of PageRank specifically aimed at images. Called VisualRank, the technology was so far only applied to a smaller test set of images, as apparently applying it to all images Google indexed would be too computing-intensive (even arguably the world’s largest super-computer can’t do everything imaginable yet). According to the New York Times yesterday, visual rank is an algorithm “for blending image-recognition software methods with techniques for weighting and ranking images that look most similar,” and in Google’s internal scoring tests it achieved far higher quality results.
If I understand the gist of the research paper [PDF] right, then it seems the core of Google’s VisualRank algo consists of not only looking at textual cues in regards to images, but also image content itself. After identifying the most authoritative set of picture candidates for a given query, Google then improves the ranking of images found to be sharing the most visual characteristics with the group at large, by creating a similarity network (which also would understand e.g. imagery shown from different perspectives, to a certain extent). Center node images or those images containing large resolution versions would then determined to be the most relevant. In 1000 sample queries – taken from the top Google Product search queries – 762 VisualRank results were tested to be more relevant than Google’s old approach, with 202 equal quality results and only 70 results that were worse.
(On a side-note, I wonder what motivates Google to publicize this information, as it could tip of their competition? Are they only being nice, potentially attracting more good researchers, or is there more to it?)
As far as I can tell, the paper does not yet indicate that Google is any further in specific image recognition, e.g. figuring out that the image found on the web contains, say, a vase of flowers without looking at textual descriptions in the vicinity of the embedded pic. 4 years ago, Google co-founder Sergey Brin said, “I don’t think that in the near future we’re going to have a service that takes a picture, and the computer decides, oh, that’s an elephant, so we search for an elephant. That seems funny to us. We should be able to do it.” Google does have face recognition features for Google Images, though; it’s found in the advanced image search dialog and works very well, and you can also use it on your own site, if you have one with indexed images, by searching Google Images for site:yourdomain.com with the face search activated.