/* ---- Google Analytics Code Below */

Tuesday, March 03, 2009

Useful Sketch Tagging Examined

I recently mentioned the service called Sketchory, which contains 250K free-to-use sketches. Like the image on the right which I got when I asked for 'abstract'. Nice service. I have been involved with the problem of tagging images. In particular choosing among a number of image types. In Sketchory I noticed that it was difficult to get to useful sketches. They were asking for aid in helping tag. So I played along, tagging a few dozen images.

It's clear after doing this why it is hard to classify. People mostly sketch images of people and animals. Sometimes fanciful things, but rarely abstract ideas and even more rarely processes that I would like to have an image of. At the right and below is an image that was retrieved, of only about a half dozen, with the term 'writing'. Even people who are trying to do this kind of tagging usefully do not do it very well. I quickly started to use the simplest nouns to tag the sketches. These, like the tags that they have already are not very useful.

So there are 250K sketches, but it is hard to find the right ones.

It is still very hard to solve the automated general tagging problem. An AI problem. A number of attempts have been made at creating tags with 'Human Intelligence Tasks' (HITs), most notably with Amazon's Mechanical Turk. It is probably better to include some prefabricated classification capability that describes a number of useful classifications. Then have the person search among many examples. Google also has an image labeling game.

See Gene Smith's good book on the subject: Tagging.


Philipp Lenssen said...

Hi, Sketchory here, thanks for the post & feedback! For some initial, pre-launch tagging we also asked at the (paid) service Uclue.com to help with the tagging. Right now, we're hoping to get more people on-board to help with the tagging. Sometimes, the crowd mind can help with problems like "what is the perfect set of tags for an image?" because once, say, 100 tags for a single image would be stored, these tags would be very varied and all-encompassing, and could also be sorted by quality (i.e. the number of repeats, like when 15 people use the tag "lamp" but only 1 person used the misspelled tag "lampp"). Mechanical Turk is also an interesting idea (which I've currently explored with CoverBrowser.com, which, by the way, also has a tagger game).

And thanks for your tagging!

Franz Dill said...

Thanks for the comments, yes good points. What I was trying to say was that different kinds of tags work for different purposes. Most of the sketches are cartoonlike of peoples and animals, not so different from typical snapshots. Appreciate your great work, will use when I can. Also check out the Google link from Louis Ahn's gamelike work. I know others have used the mechanical turk for photo tagging.