December 30, 2010

  • Rorschach's Goggles

    Today I wandered around the house with my smart phone in hand and submitted random pictures to the Google Goggles application. The application analyzes an image, searches Google's image database, and then Google's servers return a list of images that it determines match your original submission.

    When you submit images that are very unique, such as a brand's logo (I used Heineken and Coca-Cola as tests), the results are pretty stunning. However, if you move in to the recognition of every day objects, the computer fails in some very abstract ways, and it is easy to see that image recognition in this context is still at in its infancy.

    I use "fail" in the sense that the search results don't seem to be accurate at initial glance. Since I've always been interested in object recognition, I decided to use the "failed" search results to determine how the application "sees". Basically, when an image is reduced to about 150 by 150 pixels, Google Goggle's has the same eye sight as I do without my glasses (-4.25). I tested this in a couple of ways, and in the end I found it was very easy to recreate the effect in Photoshop.

    Below are several sets of images. The first column of each image is the exact picture I submitted. The additional columns are samples from the returned search results (note that the search results never returned an image anywhere close to the original object I photographed). Next I duplicated each row. Then I blurred the second row until it looked the same to me with my glasses on as the first row looked with my glasses off.

    Though the results seem abstract, you can quickly see how Goggle's computers determined a potential match. The computer has blurry vision and doesn't even begin to attempt object recognition.

    (These images have all been reduced to fit on the page)


    Truck, Star (with astronomical information), Solo dancer on stage (yes, that is one person)

     


    Salt shaker, Child in a dress, Firetruck!

     


    Toy crane, People climbing over a fence, Long jumper, Abstract painting

     


    Bananas, Dog, Portrait of a nude. On an interesting note, this also returned an image of a naked stripper. However, the image was not available online, but when blurry she seriously looked exactly like the blurry bananas.


    Apples on a tray, Samurai painting, Crappy Sega character (?)

Comments (3)

  • THAT IS FUCKING LU BU

    DYNASTY WARRIORS

    YOU GRASSHOPPER

  • For the most part, it got the concentrated colors in the right spot, although the third one not so much.  Fascinating.
    It seems like the type of thing which would inspire a photography artist, though -- a series of completely unrelated images echoing each other in form.  Heck, I may even attempt something with that concept.

  • @kirbym - Forgive me! I only played Dynasty Warriors once -- a demo in a store.

    @lanney - It is really fun to play with. I need to wander around and find some more examples. If you do some art on it, be sure to post about it.

Comments are closed.

Post a Comment