Dropbox has made some pretty smart (and frugal) acquisitions over the past couple years (including Monday’s acquisition of PiCloud) as it attempts to grow into a full-on platform, but its coolest might also be the one that very few people seem to have noticed: Anchovi Labs. It happened in September 2012 and brought some serious computer vision brains into the Dropbox fold.
Admittedly, the Anchovi acquisition slipped under my radar too, until I heard a couple recent references from people in the know about the computer vision team Dropbox is putting together. Anchovi was founded by four computer vision researchers — all Ph.Ds. — two of whom, Boris Babenko and Peter Welinder, are now engineers at Dropbox (according to the company’s personnel page). The other two, Serge Belongie and Pietro Perona, are professors at Cornell and Caltech, respectively. Belongie’s LinkedIn profile lists him as a Dropbox consultant.
This YouTube video of Anchovi’s erstwhile service shows how it worked, a process by which humans trained Anchovi’s artificial intelligence algorithms on the images they want classified (e.g., swimming pools or cats). Once the models have the training data, they can go through an entire collection of photos — and any more that might come in — and annotate them based on what they contain.
[youtube=http://www.youtube.com/watch?v=8TmjULyQ6bU&w=640&h=360]
There’s actually a battle underway in the cloud photo-storage space right now between Google, Yahoo and Facebook, and artificial intelligence is playing a big role. Google famously acquired DNNresearch, the deep neural network (or deep learning) startup of University of Toronto deep learning researcher Geoff Hinton (who’s now a part-time researcher at Google), and has rolled out auto-tagging of images in Google+. Flickr parent company Yahoo has acquired two computer vision startups in the past few months — IQ Engines and LookFlow — and Facebook is building its own deep learning team that could help with image recognition on its huge cache of user photos (although it has been tagging faces in images for quite a while).
Anchovi’s underlying techniques weren’t based on deep learning (this 2008 paper co-authored by three of the four founders seems to explain its classification model in more detail), but they still appear rather effective.
Should Dropbox want to enter the fray and start doing something intelligent with the petabytes of untagged photos that its users might some day want to search through, the Anchovi team’s skillset would certainly come in handy. It would also seem to make a nice complement for the talents of the team that built SnapJoy — a photo-aggregation service that Dropbox acquired in December 2012 before closing it in June.
I have reached out to Dropbox for a comment on the Anchovi acquisition and any plans to launch a photo service, and will update this post when I receive a response.
