Google’s Orkut filters sexually-explicit photos
28 10 2008By Joshua Koopferstock
The folks over at the Orkut blog announced a few days ago a new filter for their social network that can screen out “sexually suggestive phrases or images”. However, no detail has been published (to my knowledge) on exactly how they filter out these images, whether they use text-based clues or somehow incorporate computer vision to detect the content of photos. Until they respond to my e-mail I cannot say for sure, but my guess is that they are using only a text-based filter like the one used for Google Images SafeSearch. If this is the case, I doubt it will be as effective on a social network where there are far fewer textual clues to put an image in context than there are on, say, a pornography web site.
Regardless of whether Orkut is using it or not, it is only a matter of time before computer vision techniques drive the next wave of “safe content” filtering. As we discussed a few weeks ago, VideoSurf, the video search engine, has already introduced an image/video-based filter. Once one major player in search adopts computer vision for content filtering, their competitors will have a strong incentive to follow or risk being seen as the “unsafe” search engine. Whether or not Google started the trend with Orkut this week, it’s only a matter of time until this application of computer vision hits the mainstream.
Filtering out sexually explicit photos sounds like an interesting challenge and problem. I wonder if it can detect the difference between a photo of a beautiful girl walking down the beach in a bikini vs. a photo of a stripper about to remove the last articles of clothing on stage? Fascinating problem.
A hard problem indeed, and one for which it is unlikely that computer vision will ever solve with 100% accuracy, if for no other reason than people in general are unable to fully agree on what content should be filtered and what is ok (comments on our Polar Rose article a few weeks ago confirm this).
That being said, a computer vision-based filter could be created that would work in enough cases (or, perhaps, in enough of the more “extreme” cases) that it would prove worthwhile to implement for a company like Google.