Visual recognition is one of the significant challenges facing search as the scale of multimedia content on the web grows. It is relatively straightforward to extract text and even audio, but the variables involved in visual recognition mean that a human eye remains the best judge of a successful search. The likes of Blinkx, Riya and eVision are trying to change that but it remains early days.
Interesting then to see the Retrievr experimental service. It allows you to draw a Paint-style sketch or upload an image and then to search a selected group of ‘most interesting’ Flickr images for matches. Rather than object recognition, e.g. find ‘chairs’, it works by searching for related blocks of colour and overall shapes (see pretty accurate results for a beach below).
Although the existing link is just something to play with at this stage and the results are mixed, the technology has powerful potential to retrieve better search results. For example, rather than just a search for a dog, you could use this in conjunction with existing search to find a dog on a blue background that fills the majority of the screen.
This could have all kinds of potential applications, such as better tagging of visual archives, finding unauthorised use of your copyrighted images, finding a particular video clip.