KWPro.net

I saw the new release of Google's Similar Image Search on a few spots earlier and checked it out myself.  This is fairly impressive as a feat.  According to TechCrunch, they're utilizing the meta data collected to find the similarities rather than actual optical recognition, so I'm at how they're able to gather images with a fair amount of certainty.

Of course, this isn't perfect.  You'll definitely see odd pictures here and there meshed with the ones you're looking for.  What is impressive is seeing some pictures showing up from different angles.  A good example is when I looked for Norika Fujiwara.  In this search, the first five came up pretty accurately in terms of showing different angles for a similar shot.  Equally impressive was seeing how these shots were derived from different websites.

Unfortunately, if you look at the remaining images, you'll see that some have absolutely nothing to do with Norika Fujiwara.  In fact, you can see that Japanese Razor Ramon character.  I have no idea how that image correlates to my Norika Fujiwara image search.

Also, you'll see flaws for anything that isn't common in the basic search results.  For instance, try searching for Reina Miyauchi.  You'll get maybe at best one or two accurate results without hitting the "similar" link.  It's surprising that the basic search results coming from the image searches are so poor compared to a regular image search.  It's as if they're getting a subset of the results that they might've indexed with the remote possibility of a similar search.  Well, it's kinda pointless if only a small subset might get lucky in having a few similar search links.

I'm guessing that they have a limited number of images indexed for the similar search results at this point.  Nice try but it looks like they'll need to put quite a bit of effort to get this baby up to a decent level.