Let's say you see this pretty face in a crowd, in a newspaper or a magazine, and want to know more about who the person is. You scan her image, crop it to just her face, go to a visual search engine and hit ‘Enter'.
Within seconds, you know everything about her. You get more pictures of her, some of them from her Facebook account and almost instantly, you've put a name to her face.
Does this excite you or scare you?
If you've been using Picasa or Facebook over the last few months, you would've realised how Picasa automatically identifies faces and asks you to tag them. If you've been using digital cameras, you would know how the machine knows to identify a human in the frame and throws just the right amount of light.
Artificial intelligence applied to face recognition may just change the Internet and the way we live our lives, as the day is not far off when anyone can find out everything about you with the help of just a snapshot.
So where exactly has technology reached with face recognition and visual search?
Ok, you can breathe. It is still in a very nascent stage and there's nothing to be alarmed about at the moment. But yes, you had better start worrying and we had better have some laws protecting privacy in place before it's too late.
Google's own Image search engine still continues to search for images primarily on the basis of keywords and the advanced options include further refining them on the basis of colour, size and format. So if your pictures are tagged by your name, it should be pretty easy for anyone to find your photographs on any social networking site or blog.
Visual search engines
There are a few beta visual search engines such as IdeeInc and Pictriev that have been toying with their face recognition applications. When we ran a few tests, this is what we found.
IdeeInc uses its Pixilimar technology and its “sophisticated algorithms” to let you find images on the basis of “colour, shape, texture, luminosity, complexity, objects and regions”. See Picture 1 to see what results it returns when we scanned an image of Shah Rukh Khan. The images are certainly of the same colour and texture and also seemed to identify elements in the picture (like the skies from our test picture) but fails to recognise the face known by millions.
Pictriev actually surprised us a lot more. It could identify Shah Rukh Khan from at least a few other pictures from other websites and films from the Internet. But to see how far it is from accuracy, take a look at Picture 2 in the box. Pictriev suggests that you upload a photograph of the frontal face, with at least an 80 pixel wide gap between the eyes.
Obstacles
So what are the biggest obstacles for facial image search?
As Vasan Sowriraja, who has been working on face recognition for developing Humecons, (emoticons with your own image), observes, “Lighting is not a constant, the angle with which people take pictures is never the same and people have different expressions. Most of facial recognition is done on the basis of retina identification but the logistics are such that Facial Image Search is going to become a field of its own. With people taking high-resolution pictures, mining for images is becoming expensive because you have to search from millions of high-resolution images. Imagine, if every image is 2 MB and the results have to be given in a few seconds… it's almost impossible now.”
Photographer G. Venket Ram has a more basic answer. “It's impossible for people to look the same when you are clicking pictures impromptu with friends, say, at a beach or a party... Every time, in terms of light and mood, the face changes. Not just in terms of time, in terms of expression… be it a smile, a frown or a grin… it keeps changing. I've known so many shots where people can't even recognise well-known actors because of the make-up, the hair and even the clothes. Women, especially, can look drastically different. Look at any model's portfolio. That's the magic of a makeover.”
