12 comments

[ 3.6 ms ] story [ 31.2 ms ] thread
Made this over a handful of nights and weekends. To use, take a screenshot of the Draw Something drawing you are trying to guess and import it into the app. That's all!
This is pretty awesome! Mind sharing how the recognition works?
It's using OpenCV's template matching function. I tried to use other OCR libraries but OpenCV ended up yielding the most accurate results.
How did fit everything in a 3M IPA? Is anything downloaded after the app is installed?
I ended up doing the processing server side - performance on older devices was shockingly slow.
How does it work? Looks very cool.
I'd be impressed if it used computer vision. I'd be even more impressed if it used Amazon's Mechanical Turk. :)
Disclaimer: I don't have the app, so I don't know for sure...

I think the easiest way to solve this problem is to read the letters available, figure out how many letter the solution is, then compare these to the database of DrawSomething words. Many times you'll get a handful of possible solutions (especially for shorter words), but often enough, there will only be one possibility.