Google Image Swirl

» see image-swirl.googlelabs.com as it lived, on the wayback machine

~ the eulogy ~

I did not find Image Swirl while it was alive. It surfaced for me years later, going down a list of things Google had switched off, and the honest position is that i am reconstructing a product i never opened. That is worth saying plainly, because the temptation with a Labs experiment is to pretend everyone tried it. Almost nobody did. That was the point of the shelf it sat on.

It went up on 17 November 2009. Type a word into it and instead of a grid of results in rank order you got a small cluster diagram, each node a thumbnail standing for a group of pictures that resembled each other. Click a node and it expanded into its own sub-cluster. The example Google kept using was jaguar, where the animal and the car separate on their own without anyone tagging anything, and Eiffel Tower, where the split is between the daytime postcard shot and the lit-up night version.

The machinery under it was a similarity graph. Google Research built pairwise scores between images out of visual features, landmark recognition, face recognition and the text around them, then ran graph analysis over the whole set to find groupings that no single pairwise comparison would have shown. Yushi Jing and Henry Rowley wrote the launch post. The lineage runs back through their earlier work on ranking images by what is in them rather than by who links to them.

~ the dossier ~

~ what it was actually for ~

Image search has a specific failure that ordinary search does not. A ranked list assumes there is one best answer and the rest are worse versions of it, which is fine for a question and wrong for a picture. When your query has two meanings, ranking just interleaves them and you scroll past one to reach the other. Swirl's answer was to stop ranking and start grouping, and to show you the groups before you had to commit to one.

The reason this stayed an experiment is that it asks the user to do a second thing. A grid of images needs no instruction. A cluster diagram needs you to understand that the node is a stand-in for a set, and that clicking it goes deeper rather than sideways. Google was in the middle of stripping steps out of everything it shipped that year. A product that adds one is not going to survive that meeting.

~ the postmortem ~

On 20 July 2011 Google announced it was winding down Labs to concentrate its people on fewer products. It was a bulk decision rather than a judgment on any single experiment, and Swirl went out with Fast Flip and the rest of the shelf. The pages lingered for a while afterward the way abandoned Google URLs do, resolving to something for months after the announcement that killed them.

Labs itself is the interesting corpse here. For about five years Google ran a public place where research could be shipped to real users without being a product, with no support promise and no expectation of survival. Closing it did not stop the experiments, it moved them behind the login where the failures are invisible. You no longer get to watch a search company try an idea in public and drop it.

~ what it left ~

The grouping idea did not die with the interface. Modern image search puts a strip of refinement chips across the top that do roughly the same job in a much duller shape, and reverse image search will happily tell you which visual cluster your picture belongs to. What went is the diagram, the part that let you see how many different things your word meant before you picked one.

Twenty months is short even for a Labs project, and Swirl never got the chance to fail on its own merits. It was closed because the building it lived in was closed. That is a normal way for research to end and it leaves nothing to argue with.

@}-->--

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

« previous  ·  random  ·  next »