GazoPa

» see gazopa.com as it lived, on the wayback machine

~ the obituary ~

GazoPa launched on 10 September 2008, on stage at TechCrunch50, and it was a Hitachi project rather than a startup. The pitch was content-based image retrieval in public and for free: hand it a picture, get back pictures that look like that picture, worked out from the image itself instead of the words printed around it.

Open beta followed in 2009. At its height the site drew more than 90,000 unique visitors a month, which is small for a search engine and respectable for a corporate research lab shipping something at consumers. A companion site called GazoPa Bloom arrived in 2010, a flower photo community where you uploaded a bloom and the engine and the crowd between them tried to name it.

I work on computer vision now, on edge hardware, so this is the direct ancestor of the problem I look at most days. The honest description of what GazoPa was doing in 2008 is feature extraction and nearest-neighbor lookup over hand-engineered descriptors, about four years ahead of the deep learning turn that made any of it work properly. It was the right question asked with the wrong decade's tools.

~ the case file ~

~ the way you searched ~

You could feed it an uploaded image or a keyword, click any result to promote that result into the new key image, or open a drawing tool and sketch the thing you were after. The sketch input is the part people remember. It was awkward and it mostly produced nonsense, but in 2009 a search box you could draw into did not exist anywhere else you could reach from a browser.

Contemporary reviews were blunt about everything else. Keyword search came back worse than Cuil's, and running the same query twice could hand you two completely different result sets. Links often pointed at a site's homepage instead of the page actually holding the image. The index sat around fifty million pictures, roughly five percent of Google's.

~ the pivot ~

The shutdown notice went up on the GazoPa blog on 31 May 2011 and named 8 June as the end of consumer service. The reason given was a move to business customers, "especially in China where e-commerce grows dramatically," which in 2011 was a sound read on where visual search would first pay for itself. Hitachi promised existing users a replacement research tool in the near future. GazoPa Bloom went out at the same time.

Some records date the closure to 23 June instead, and those two weeks are the space between the announced end of service and the site actually going dark. gazopa.com now resolves to a parking page on a numbered subdomain, which is the ordinary fate of a lapsed corporate domain and says nothing about what used to be there.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »