Pages

Sunday, March 5, 2006

Capture v. Derive

Universal Law: It is easier, cheaper and more accurate to capture metadata upstream, than to reverse engineer it downstream.

Back at Virage, we worked on the problem of indexing rich media - deriving metadata from video. We would apply all kinds of fancy (and fuzzy) technology like speech recognition, automatic scene change detection, face recognition, etc. to commercial broadcast video so that you could later perform a query like, "Find me archival footage where George Bush utters the terms 'Iraq' and 'weapons of mass destruction.'"

What was fascinating (and frustrating) about this endeavor is that we were applying a lot of computationally expensive and error-prone techniques to reverse engineer metadata that by all rights shoulda and coulda been easily married to the media further upstream. Partly this was due to the fact that analog television signal in the US is based on a standard that is more than 50 years old. There's no convenient place to put interesting metadata (although we did some very interesting projects stuffing metadata and even entire websites in the vertical blanking interval of the signal.) Even as the industry migrates to digital formats (MPEG2), the data in the stream generally is what is minimally needed to reconstitute the signal and nothing more. MPEG4 and MPEG7 at least pay homage to metadata by having representations built into the standard.

Applying speech recognition to derive a searchable transcript seems bass-ackwards since for much video of interest the protagonists are reading material that is already in digital form (whether from a teleprompter or a script.) So much metadata is needlessly thrown away in the production process.

In particular, cameras should populate the stream with all of the easy stuff, including:

  • roll
  • pitch
  • yaw
  • altitude
  • location
  • time
  • focal length
  • aperture setting
  • gain / white balance settings
  • temperature
  • barometric pressure
  • heartrate and galvanic skin response of the camera operator
  • etc.
  • Heartrate and galvanic skin response of the camera operator? Ok, maybe not... I'm making a point. That point is that it is relatively easy and cheap to use sensors to capture these kinds of things in the moment... but difficult (and in the case of barometric pressure) impossible to derive them post facto. Why would you want to know this stuff? I'll be the first to confess that I don't know... but that's not the point IMHO. It's so easy and cheap to capture these, and so expensive and error-prone to derive them that we should simply do the former when practical.


    An admittedly slightly off-point example... When the Monika Lewinsky story broke, the archival shot of her and Clinton hugging suddenly became newsworthy. Until that moment she was just one of tens of thousands of bystanders amongst thousands of hours of archival footage. Point being - you don't always know what's important at time of capture.

    So segueing to today... Marc, Ellen, Mor and the rest of the team at Yahoo Research Berkeley have recently released ZoneTag. One of the things that ZoneTag does is take advantage of context. I carry around a Treo 650 with Good software installed for email, calendar, contact sync'ing. When I snap a photo the device knows a lot of context automagically, such as: who I am, time (via the clock), where I am supposed to be (via the calendar), where I actually am (via the nearest cell phone tower's ID), who I am supposed to be with (via calendar), what people / devices might be around me (via bluetooth co-presence), etc. Generally most of this valuable context is lost when I upload an image to Flickr via the email gateway. I end up with a raw JPG (in the case of the Treo even the EXIF fields are empty.)

    ZoneTag lays the foundation for fixing this and leveraging this information.

    It also dabbles in the next level of transformation from signal to knowledge. Knowing the location of the closest cell phone tower ID gives us course location, but it's not in a form that's particularly useful. Something like a ZIP code, a city name, or a lat/long would be a much more conventional and useful representation. So in order to make that transformation, ZoneTag relies on people to build up the necessary look-up tables.

    This is subtle, but cool. Whereas I've been talking about capturing raw signal from sensors, once we add people (and especially many people) to the mix we can do more interesting things. To foreshadow the kinds of things coming...

    • If a large sample of photos coming from a particular location have the following tag sets [eiffel tower, emily], [eiffel tower, john, vacation], [eiffel tower, lisette], we can do tag-factoring across a large data set to tease out 'eiffel tower.'
    • Statistically, the tag 'sunset' tends to apply to photos taken at a particular time each day.
    • When we've got 1000s of Flickr users at an event like Live8 and we see an upload spike clustered around a specific place and time (i.e. Berlin at 7:57pm) that likely means something interesting happened at that moment (maybe Green Day took the stage.)

    All of the above examples lead to extrapolations that are "fuzzy." Just as my clustering example might have problems with people "eating turkey in Turkey", it's one thing to have the knowledge - it's another to know how to use it in ways that provide value back to users. This is an area where we need to tread lightly, and is worth of another post (and probably in fact a tome to be written by someone much more cleverer than me.)

    Even as I remain optimistic that we'll eventually solve the generalized computer vision problem ("Computer - what's in this picture?"), I wonder how much value it will ultimately deliver. In addition to what's in the picture, I want to know if it's funny, ironic, or interesting. Much of the metadata people most care about is not likely to be algorithmically derived against the signal in isolation. Acoustic analysis of music (beats per minute, etc.) tends to be a poor predictor of taste, while collaborative filtering ("People who liked that, also liked this...") tends to work better.

    Again - all of this resonates nicely with the "people plus machines" philosophy captured in the "Better Search through People" mantra. Smart sensors, cutting-edge technology, algorithms, etc. are interspersed throughout these systems, not just at one end or the other. There are plenty of worthwhile problems to spend our computrons on, without burdening the poor machines with the task of reinventing the metadata we left by the side of the road...