Monday, May 26, 2008

Labeling images for Google

I was given a link the other day to Google's new image labeler: http://images.google.com/imagelabeler/

What is interesting about it is that everything is done in pairs and then the pair is given images to label and points for how many labels they can come up with, more points for more specific labels.

Then what do you do with all those points? What do you get from being on Google's website and say that you have "24300260 points"? Not much, but there is a lot of evidence out that that some people might just have too much time in their hands. Better to ask them to do these things than them being out and about playing with friends, don't you agree?

Anyway, that's not the reason why I've decided to post this here. What I found interesting is the names of the people that I saw on the "today's top pairs" section. I'm not too worried about pasting them here because they are quite anonymous:

1. Wonderful Now 3K - mY record 2980
2. BSPECIFIC - Me 2 GottaGoNow 2880
3. Me 2 GottaGoNow - Lds Labelmaker 2620
4. guest - guest 1950
5. Monster - DoubleDogsRacing 1790

They seem like people use names to communicate their intentions with other labelers. Like "BSPECIFIC" or "Wonderful Now 3K". Also there are some that look like an automated system: "Lds Labelmaker", which might relate to the Low Density
Separation algorithm [1] or maybe Limited Discrepancy Search [2], or maybe it's somebody from the Church of Jesus Christ of Latter-day Saints... Who knows?

The fact is that eventually Google will figure it out and decide if their little points-giving experiment will help or hurt their image search. I bet on the latter. I actually bet that just adding humans to the loop will not really help you in general. but might provide enough test data to then add an automated method and have an idea of how well it's doing.

[1] - Chapelle O., Zien A., (2005). Semi-Supervised Classification by Low Density Separation, 10th International Workshop on Artificial Intelligence and Statistics, AI STATS 2005, Barbados]
[2] - L.D. Whitley, A.E. Howe, S. Rana, J.-P. Watson, and L. Barbulescu. (1998) "Comparing Heuristic Search Methods and Genetic Algorithms for Warehouse Scheduling." In Systems, Man and Cybernetics, 1998.

0 comments: