Flickr is probably the greatest online photo management and sharing application to hit the internet. You can spend hours browsing through millions of beautiful images from around the world, which are categorised and tagged for your convenience. Recently, Christian Langreiter released a new experiment named Retrievr. Retrievr allows you to sketch something in your browser and it when then search the Flickr photo database for images which it thinks matches your sketch!

Retrievr is based on research conducted at the University of Washington on a topic called Fast Multiresolution Image Querying. To describe the process simply, you take an image and create a wavelet of the image. Using a wavelet transform, you could generate many different representations of the same base data – a lot like varying the compression level when saving an image. From these wavelets, a signature of the image is formed composed of the key wavelets, while all non-significant items are discarded. Once a wavelet has been generated for each image, they are stored in the database for fast retrieval later. As a user of Retrievr, you simply create a sketch, it computes the wavelet for your sketch, compares it against the wavelets already stored in the database and returns you a set of best matches. At this stage, only a small subset of the images on Flickr have been analysed for use in Retrievr, however Chris says to email him if you’d like to see another group/set of images included into the site.

Ultimately, I think it is a awesome experiment which proves what is possible through utilising Fast Multiresolution Image Querying however at this stage I can’t can (see comments below) see a real world practical use for it.

3 thoughts on “Retrievr

  1. I love these ‘proof of concept’ things. They are a necessary step in communicating the value of this kind of technology to those who might choose to invest in it.

    Sites which might benefit most from this sort of thing in the short term are Stock Photo providers like — as they already maintain their own DB of images and while their stock is large, their uptake of new material isn’t so staggering that they can’t reasonably generate ‘wavelets’


  2. Imagine being able to search on a stock photo site for images matching a particular set of palettes or tones. You could put sky blue on the top and and grass green on the bottom and see what photos they might have on offer to suit what you’re looking for. That’d be a very useful feature I think.


  3. yeah, tonal matching would be great!… even something as simple as pulling up photos with a high concentration of colour similar to your corporate colour.

    There’s definitely a demonstrable use for this. The more time people can save when looking for stuff like this, the more you can charge for the servcie.

Comments are closed.