Automatic Image Recognition is still a Marketing Pitch

Automatic image recognition sounds very impressive. Sometimes it also looks very impressive. But it has no future when it comes to describing image content.

Take Riya, a California-based start-up that launched a photo service promising to recognize faces on your photos. It did sound very impressive (like everything that includes the word “automatic”) when you were said that you could upload 1,000 photos, click a magic button, and the system will recognize your face on all of them. However, when they reached objective technology limitations, they decided to re-think the approach and launched Like.com.

It’s the Web site that allows you to upload a photo which contains a product, for example, President Bush’s watch, and then find another watch looking the same or similar in Web shops. However, as image recognition technology still has limitations and successful recognition depends on multiple factors, visual search on Like.com is currently limited by objects that can be easier identified: handbags, watch, jewelry,and shoes.

It means there is a long way from a marketing pitch to the system that really works. Robert Scoble who interviewed Munjal Shah, Riya’s CEO and co-founder, a few days before Like.com has been officially launched, says in his blog, “Riya was pretty close to being sold to Google. If it had been, they never would have worked on this search engine. So, by getting turned down by Google Riya came back with a much better business.”

Well, probably it's not too bad. Munjal Shah said that woman fashion market worth USD 15-30B so targeting this specific niche can be a good business.

However, what about image search for everyone? Can automatic image recognition help recognize individual elements of image content and index them for better search? There are still things when the human brain does not have a replacement. I doubt Riya or any other image recognition system can be really useful when it comes to describing what is shown on my photos from Budapest and telling facts I know about people and places on the pictures. Adding a semantic layer, which can be used then for better indexing, is not about automatic image recognition. It's all about people.

Tags: Chicago, John Hancock Center, 900 N. Michigan, Olympia Center, Chicago Palace, City Palace, Marina Towers, 77 W.Wacker

I believe social-based image tagging is a much more promising approach. By the way, Google and Yahoo are making efforts that demonstrate they would rather motivate users to manually describe images than relying on automatic recognition results. The Google Image Labeler game is a good example, though I personally doubt it is able to provide high quality tagging. After all, if a photo contains multiple objects and you describe each of them by adding a tag, how one who found this photo can correlate the tags with the actual content? Have a look at the photo on the left to see what I mean. Now can you tell me where is the City Palace here? Probably, you can't unless you live in Chicago. What we are doing with our FotoTagger image annotation technology already implemented in a series of tools is going down to the lower level of granularity and allowing a user to connect visual content to textual description.

Take a look at the same picture. Now it is annotated by FotoTagger. It looks much clearer, doesn’t it? By the way, the tags can be hidden in a click so you can see the "pure" original picture.

This object-specific description annotation is stored as an XML description embedded into an ordinary JPEG file. It means everyone can win here: search engine companies, Web site owners, and Web surfers.

Here you can read an article on how we offer to meet this challenge.

Posted on December 11, 2006 by Alex Masycheff

See all posts