How do I find Waldo with Mathematica?
stackoverflow.com
stackoverflow.com
Another commenter got me curious about just how hand-tuned the answer was, so I fired up Mathematica 8 and put their solution in. I had to dig up the original image using Wayback Machine (thanks, archive.org!) but then, to my surprise, their answer didn't actually work. It merely darkened the image uniformly, meaning it wasn't able to find any potential Waldos.
I fiddled around with their solution for a bit, and then finally just hooked their parameter up to a slider. I thought maybe the "level" parameter just needed a bit of tweaking. But surprisingly, it still doesn't work as advertised: http://i.imgur.com/XWNTvYU.jpg
Their "level" parameter was originally 0.12, but like I said, that didn't work at all. It doesn't work until the parameter is bumped up to 0.18, at which point it gives the wrong solution (the lower-left circle). Then, 0.19 gives another incorrect solution (the lower-right circle). Finally, 0.20 identifies Waldo.
So this solution is a nice way to filter out potential targets, but I wish I could figure out whether the original poster was being shady or whether Mathematica's internals changed since they posted their answer. I'm inclined to give them the benefit of the doubt, but that has me burning with curiosity about what exactly must've changed in Mathematica's implementation.
EDIT: I didn't do the most basic of checks: to make sure I didn't introduce any changes in the original experiment. It turns out by rehosting the image on imgur, it compressed the resulting image, introducing artifacts into the color channel. Their solution no longer worked as advertised due to this. That shows how brittle the solution is, though! I wonder if we can make it more robust without fundamentally rewriting it?
UPDATE: Changed the correlation value from .12 to .2 and the compressed imgur image has some interesting results (3 circles, one of them is on Waldo): http://cl.ly/image/0Z3k150S2z10
Yeah, I'm getting the exact same results. So it looks like compression is messing with the algorithm's correlation ability. I think the problem is that red color values are often thrown out in image compression (moreso than intensity / green channels).
return (150, 200);
Obviously I'm being pedantic, and I mean no disrespect, but I have a bone to pick with so-called "computer vision" algorithms that are little more than simple image processing. In this case, the time spent implementing and tuning the algorithm exceeds the time it would take to solve the task manually (e.g. pay a second-grader to do it). And as others have observed, it isn't obvious that this algorithm would generalize to other images, in which case no time is saved over the manual approach.
It is tempting to dismiss sophisticated techniques because (a) they are hard to understand and (b) the task seems so easy to our own brain-equipped vision systems. But the fact is, most interesting computer vision problems (including this one) require sophisticated representations to achieve robustness and generality. In other words, any good solution will need an answer to the question "What is a Waldo?" that is better than "a 50x50 patch of pixels with red and white stripes".
Implementing any learning algorithm should be the absolute last resort. They're complicated, hard to generalize, and difficult to guarantee in the real world. Not that image processing is much better, but it's definitely more deterministic and predictable in its results. With any system geared toward a specific goal, your first approach should be to look at the simplest solutions (no matter how rudimentary they may be).
> It is tempting to dismiss sophisticated techniques because (a) they are hard to understand and (b) the task seems so easy to our own brain-equipped vision systems. But the fact is, most interesting computer vision problems (including this one) require sophisticated representations to achieve robustness and generality. In other words, any good solution will need an answer to the question "What is a Waldo?" that is better than "a 50x50 patch of pixels with red and white stripes".
Yes, of course. I'm using waldo as an example toward a problem I often see reading research papers, that is an overly complex system made to try to solve more than just the initial task. "What is a Waldo" is irrelevant to the task at hand. Identifying the best candidate for Waldo is the goal. In this situation simple correlation (or even more efficient correlation methods: MMCF, OTSDF, QCF, etc) can give good results (except on that all waldos page...damn that's even hard for a human) most of the time.
To make a euchre analogy - you don't pass on a biddable hand, and you don't waste your trump cards.
"53 percent of the time Waldo is hiding within one of two 1.5-inch tall bands, one starting three inches from the bottom of the page and another one starting seven inches from the bottom, stretching across the spread."
http://www.slate.com/articles/arts/culturebox/2013/11/where_...
If you repeated this with random data, I expect you could find the same results: this does not constitute evidence that Waldo is placed in those bands with priority.
You can mathematically say that 100% of the time, Waldo appears in 0% of the page (defined by individual points rather than areas). This equivalent statement clearly has no predictive power.
In my opinion, it would not be reliable to use this on novel Waldo images.