But did it keep Mathematica quiet while it looked?
March 23, 2012 6:15 AM   Subscribe

How do I find Waldo with Mathematica?
posted by OmieWise (10 comments total) 13 users marked this as a favorite

 
I'm kind of surprised by the attention this is getting on the internets; this was a week 1 homework question in my image processing class about 10 years ago. Not trying to brag, as I wasn't able to do it, but most of the other people in my class could do it! I think most people did it in OpenCV or MatLab, using template matching.
posted by krunk at 6:25 AM on March 23, 2012 [1 favorite]


But what about that last page where Waldo is in a whole mob of Waldo clones and other people in red and white stripes, and the only way to find the real Waldo is because he's missing a shoe?
posted by Faint of Butt at 6:27 AM on March 23, 2012 [6 favorites]


Paging Werner Herzog

It's not a self-link if I posted it on MetaFilter, is it?
posted by Deathalicious at 6:43 AM on March 23, 2012


What krunk said. Sheesh.
posted by erniepan at 6:58 AM on March 23, 2012


Extra points for navigating the drone and delivering the Hellfire missile.
posted by RobotVoodooPower at 7:20 AM on March 23, 2012 [3 favorites]


Easy, he's at the bottom of the ocean now.

No, wait, which one was Waldo again?
posted by Cironian at 9:16 AM on March 23, 2012


yea thats a part of image recognition system. you have taken a step ahead towards expertise if you did that.
posted by johnstendicom at 9:18 AM on March 23, 2012


I think this activates the same part of our brains in terms of "woah, the computer did this amazing thing" that this metafilter thread did. Still makes me happy, thinking of that.
posted by jscott at 12:08 PM on March 23, 2012 [1 favorite]


Being amazed by what the computer managed to do is the story of all computer vision algorithms. There's something really special on doing complicated arithmetic and getting a simple visual result. When you're developing a vision algorithm, you usually start with a single image and try to make it work. When it finally does, you can see the computer's answer is obviously correct, and you are inclined to think that the machine has succeeded in doing the same kind of process you did yourself with your brain, and it is a magical feeling.

After that, however, becomes disappointment. There's a huge difference between tuning your method so that you get the correct solution in a single image, and making it work in, say, 90% or 99% of all possible images. As you try to iron out the glitches in your ingenious and increasingly complex method, it becomes excruciatingly clear that the computer is not solving the problem in the same way after all, as it will keep tripping on input images which you yourself can solve without much conscious effort.

Then comes the lure of self-deception. You will start weeding out "outliers" from your data and redefining your problem statement if you can. You feel that you're really close to a good solution, and it's the data that is at fault, not your method. Finally, after careful tuning of all the parameters, you manage to get something that seems to be working for most of your data set. But you're probably still wrong. What you've managed to do is to the method work for the data set you have, but if you try it with other similar data sets, it does not work nearly as well. In machine learning theory, this is called overfitting, and it's very hard to avoid completely.

You fall into disbelief. How is it possible for anyone to learn anything, if it is this hard to make a computer do it well? Is all learning just some form of overfitting, making overly optimistic predictions which will fail catastrophically as the input data set changes just a little bit? You find comfort in reading (or watching) The Black Swan, once more. You're not the only one doing these mistakes, over and over again.

At this point, if you manage to take a step backwards and review your work objectively, you'll see that you have something that works in a lot of cases. The computer really does not solve the problem the same way as you, and it will perform worse than you especially on surprising input, but it might be good enough. How does the computer do it then? Although you know all the parts by heart, you can't explain it fully. It just all works together. It's a little bit of amazing.
posted by ikalliom at 4:02 PM on March 23, 2012 [2 favorites]


Ooh can we do this with probabilistic graphic models? (Is this gonna be on the quiz)
posted by jcruelty at 12:11 AM on March 24, 2012


« Older 17 Vintage Comic Book Covers Where Superman is a C...  |  The Turkana Basin Institute an... Newer »


This thread has been archived and is closed to new comments