• United States

Retrievr aims to find specific images in Flickr

Jan 25, 20063 mins
Enterprise Applications

* Draw the image you want to find on Flickr and Retrievr aims to do the rest

While the Web is effectively indexed through text, the same can’t be said about images. If you have ever looked for a specific image the chances are that unless someone happened to create a useful ALT attribute or the page it is on was seeded with useful keywords you will be out of luck.

A new service that has some promise in this area is Retrievr from SystemOne, a consulting company based in Austria.

The Retrievr Web application consists of a Flash-based sketchpad on the left and a results area on the right. You draw simple diagrams of the shapes you are looking for in the colors you want and Retrievr matches your doodles to photographs from the Flickr photo service.

Does it work? Yes, mostly. Does it work well? Sometimes.

The underlying algorithm is based on something called Wavelet Transforms (I won’t try to explain this technique as I barely understand it but if you are so inclined, check out “A Really Friendly Guide to Wavelets”). The application of wavelet-transforms to pictures is discussed in a paper titled “Fast Multiresolution Image Querying” that the developer of Retrievr based his algorithm on.

For what it is worth, Retrievr was written in Python, one of my favorite languages (see Gearheads “A high-level language worthy of your tool kit” and “Python reappears, with 100% pure Java”, not to mention “Python eats up white space”).

The way Retrievr works is to scan pictures, stuff them through the wavelet transform algorithm, and extract what could be thought of as a “fingerprint.” When you draw your model on the sketchpad it is put through the same process and its fingerprint compared to those in the database. The images associated with the best matches are then displayed.

Go on, try it! A red filled-in circle will select pictures of poppies and roses as well as more pictures with large red blobs. A series of horizontal blue lines will match to pictures of water and distant hills while a black background with a large white dot produces a picture of the moon and a variety of faces.

On the other hand, a sketch of a wine glass produces a picture of a finger with a face drawn on it and a noose around where a neck would be. Weird to say the least.

All the same, this is exciting and for Web applications, there a lot of potential particularly when you have a limited set of images to match. I’d love to see an interface that allowed you to upload photos so that you could in effect ask the system to find more of the same.

I’d be interested to hear what you make of this and whether you might use such a system in your projects.


Mark Gibbs is an author, journalist, and man of mystery. His writing for Network World is widely considered to be vastly underpaid. For more than 30 years, Gibbs has consulted, lectured, and authored numerous articles and books about networking, information technology, and the social and political issues surrounding them. His complete bio can be found at

More from this author