Flickr Search Hack Powered by Mouse-Made Doodles

Carl Bialik from WSJ writes "Retrievr gives budding artists an impractical but addictive way to find photographs on Flickr: a search engine powered exclusively by mouse-made doodles. From the article: 'Retrievr, Mr. Langreiter says, "doesn't look at specific forms." Art history buffs might like to think of it as photo-search by way of Impressionism. The Retrievr engine dissects a photo like a gallery connoisseur who lost his bifocals: It focuses on regions of colors rather than specific shapes and lines. "It is, actually, a simple scheme," says Mr. Langreiter. Retrievr creates and stores a compact representation of each photo in its database. The system pulls only the most important features — broad shapes, blocks of color and spatial relationships between different colored areas — out of detailed images to create shorthand approximations of every photo. (The storage mechanism extracts the 120 "strongest" features from an image to create something called a "wavelet transform," which contains much less data than the photo itself and facilitates lightning-fast searches.)'"

Comments

Flickr Retrievr

[tonyr1988]

Direct Link

Requires Flash.

It's been done before

[pyite]

I read about this a little while ago. Same principle. It uses a Haar transform (for those unfamiliar with multimedia signal processing and wavelets, specifically, the Haar transform is a specific wavelet transform based on the Haar wavelet and the associated orthogonal basis). The idea is that you compare the low frequency component of an image to the low frequency component of a rough drawing (which is pretty low frequency to begin with) and they should be pretty close of the images have anything in common.

Re:It's been done before

[pyite]

Haar wavelet though? While it's easier computationally (since the mother/father wavelets are peicewise linear in the 1D case) I always saw it as being a "lesser" wavelet in the sense of compression/reconstruction quality and ability to discern edges/other dramatic changes in data

I'm just saying what imgSeek uses. It's certainly a very easy wavelet to implement via lifting. I think it's probably used because more complex wavelets wouldn't be of any help since the rough drawing is so rough to begin with. In the end you could probably do the same thing with a DCT. Wish I had time to experiment.

Rating the doodles

[inKubus]

Keep in mind that there's a rating system for the doodles also.. there's some pretty cool artwork in there, as well as 50% boobies, dicks and strange V shapes (everyone draws them a little different). Pretty fun, it's under the Art of Retrivr

wavelet first

[mennucc1]

just for the record: the post claims that ``the storage mechanism extracts the 120 "strongest" features from an image to create something called a "wavelet transform"'' but this is quite misworded. Indeed if you look into the original research project, you see that ``the algorithm performs a wavelet transform on every image, and then collects just the few largest coefficients from this transform''.

O-o-old

[remmelt]

http://www.boingboing.net/2006/01/04/drawing_inter face_fo.html Still cool, but it's flopped in the Flickr Community because it's not that good at actually finding the pictures. It's more colour based than shape based.

Urg

[hyfe]

(The storage mechanism extracts the 120 "strongest" features from an image to create something called a "wavelet transform," which contains much less data than the photo itself and facilitates lightning-fast searches.)'"

If you're going to simplify, atleast get it somewhere near correct. A wavelet transform doesn't extract features. Features is a human-made concept. A wavelet transform is simply a transform (or for this purpose, a very lossy compression algorithm), very similar to the Fourier Transform, except that it has locality which is why it performs soo much better on non-uniform data.

I mean, 'something called a wavelet transform'. A short explanation linking it Fourier might have been apt, but wavelets are hardly voodoo.

'facilitates lightning-fast searches'.. oohh, thanks for telling us. I would never have guessed that after transforming the data down to 12 vectors, searching would be a lot faster. I mean, if they actually had indexed the data in a clever way or something specifically to speed up searches, this sentence would have made sense.. but they just transformed it. It's not voodoo and market-speech is bad!