Tuesday, 27 May 2008

Google image labeler and VisualRank - improving Google Image search and serving display ads

Display ads could soon be served against Google image results. This would result in a significant increase in Google's display inventory and “represents a large opportunity" according to Marissa Mayer, Google VP of Search Product and User Experience.

"Google calculated in 2006 that it was giving up as much as $200MM a year by not including text ads [against image searches.] Trials showed that text ads drove people away from conducting image searches, and Google dropped that idea. Display ads may work better with image searches because they seem more natural to people looking for pictures." (http://www.bloomberg.com/apps/news?pid=20601087&sid=a1t9ASFGIJlY)

At the recent Google AGM, monetising YouTube and growing display revenues were identified as priority areas for Google moving forward (http://www.guardian.co.uk/business/2008/may/09/google.youtube). However, to offer the same level of contextual relevance in display as Google do with text based searching is a challenge. With text queries, computers are reliable in establishing user intent and serving relevant results / ads. Interpreting images and video is significantly more challenging and Google are currently using 2 routes to help improve image and video recognition:

1) Google Image Labeler

I was recently sent a link to Google Image Labeler (http://images.google.com/imagelabeler/). This program is positioned as a sort of image recognition game, enabling users to help label Google's images and help improve the quality of Google's image search results.

Google Image Labeler users are randomly paired with a partner, served a set of images over a 2 minute period and enter labels for the images shown. Labels that match the other participant gain points with a hall of fame table published on the Google Image Labeler home page. When the 2 minutes is up, users are shown all the images that they viewed and given the web addresses so that they can further explore the websites holding these pictures.

Google Image Labeler is strangely addictive and the more people that use it, the more Google's systems will develop in terms of image interpretation. A clever way of crowdsourcing and getting Google Images users to do the donkey work!

2) Development of VisualRank - PageRank for image and video

VisualRank is an experimental program that seeks to enable computers to produce a 'PageRank for images.'

Recently announced in Beijing, VisualRank is "a system that used visual cues, instead of solely text information, to determine the rank of images. The idea was simple: find common visual themes in a set of images, and then find a small set of images that best represented those themes. The resulting algorithm wound up being PageRank, but on an entirely inferred graph of image similarities." (http://googleresearch.blogspot.com/2008/05/visualrank.html)

VisualRank has been designed to ensure that the most relevant images or video are returned against enquiries. Not yet live, it is research in progress, but expect this to become as important as PageRank - if nothing else it could kickstart a whole new branch of SEO based around image / video optimisation.

Google are therefore not sitting back and waiting for display money to roll in. The purchase of Doubleclick, the recent introduction of third party ad serving / tracking on the Google Content Network (http://googleblog.blogspot.com) and the projects outlined above all seek to expand Google's display offering. These moves are designed to prevent Google from becoming commoditised (by being seen solely as a Search provider) and seek to enhance the Google brand for both users and advertisers.

Interesting to see how Google's display capability will develop - will Google ever challenge Yahoo! in display?

No comments: