Not quite there: fun with Google Image recognition and search

There's something playful about the way researching one topic often leads to new insights on another.  While attempting to track down copyright information for a test file, I decided to try out some of Google's newer image recognition and search capabilities, such as the browser-based Google Image and the mobile app, Google Goggles.

The underlying idea is pretty wonderful – instead of trying to think up just the right search terms, simply snap a picture or upload an existing image, and the magic of Google will find images “like” it.   This could be great tool for designers who may be looking for similar but slightly different images to fit into a specific design project.  Could the promise of algorithmic image recognition reduce the need to apply keywords and descriptions to our images?

Already  having an image in hand, I turned to Google Image.  Here’s the reference image I used, a greyscale pair of figure skates.  Simple, right?


 
Here is a sampling of the delightfully "similar" Google Images results:

 
 Source: http://nawel-louerrad.blogspot.ca/
 
 Source: http://www.carid.com/db-exhaust/db-sport-polished-stainless-steel-exhaust-system.html

 Source: http://americanreviewmag.com/opinions/reach-for-the-stars

 Source: http://www.thefineprintuf.org/2012/10/23/diy-bike-mechanic/
 
 Source: http://solilustraciones.blogspot.ca/

 Source: http://www.capital.bg/light/revju/tester/2012/09/28/1915276_petdeset_njuansa_sivo/

 Source: http://www.sciencedirect.com/science/article/pii/S1470180409015168

Note, These all appeared in the top 10 results.  It seems we are not yet ready to leave behind quality descriptive metadata, such as keywords, headlines and descriptions, to ensure good search results in our digital asset management systems.  We may safely say that image recognition capabilities are anything but:

 
 Source: http://luc.devroye.org/uae.html (Search result 19, btw)

Postscript - in fairness to Google, I must note that you can improve precision significantly by adding search terms in addition to the reference image: