Search, Ambiguity, and Autotagging

At the beginning of this year, there was a burst of publicity around the development of a European public-private multimedia search engine project called Quaero. Billed as the European challenger to Google, the backers evidently didn’t like the publicity, because they took everything about the project offline. It was slated to be launched in the spring of 2006, but is still incommunicado, and there seems to be a fair degree of scepticism about the eventual success of the project.

What interests me is that this search engine uses pattern recognition to enable searching of images, audio and video files. The technique used is called “keyword propagation”. What happens is that when (say) a portion of an image is highlighted and tagged with a keyword, the engine crawls all other image content on the web, and assigns that keyword to any other similar visual patterns. Great leap forward for mankind, one might think, in doing image searches for photos of Albert Einstein, we no longer need somebody to have previously tagged that photo with his name. One of Quaero’s ambitions is to enable auto-tagging of vast quantities of audio visual archive material.

In a slightly different (and more public) context, Flickr competitor Riya is now applying similar technology to a social tagging context: you highlight your image, tag it, and hey presto, anything else in the Riya space with similar patterns will also get auto-tagged.

Now auto-tagging has been around for a while in the text world, and it’s still slightly problematic. You generally need hefty investments of expert’s time to train the engine to recognise the true subject matter of a document within the enterprise environment, otherwise the inherent ambiguities of language make the search results quite imprecise.

What I’m curious about is how this will work with images in an uncontrolled social tagging context. With sufficient volume of content and users, the social tagging phenomenon based on text has demonstrated an intriguing challenge to the taxonomist’s demand for vocabulary control. The subject ambiguity of text-based documents however is nothing when compared with the ambiguities inherent in images, not to mention more complex media such as video. In Flickr, a relative degree of control is achieved by having lots of people tag the same images.

How will ambiguity work with propagation? My guess is many less than useful words will be assigned to images by autopropagation (I can imagine “nose” “teeth” “smile” “ugly” “sweetoldman” could all be added to Albert Einstein-lookalike images).

I really want to see how Riya develops and responds to this (thanks Maish for the Riya link).

0 Comment so far

Commenting is not available in this weblog entry.

Comment Guidelines: Basic XHTML is allowed (<strong>, <em>, <a>) Line breaks and paragraphs are automatically generated. URLs are automatically converted into links.