[Sigia-l] Video metadata

[Eric Scheid]

On 11/4/06 5:52 AM, "Peter Van Dijck" <peter at poorbuthappy.com> wrote:

>> I always use the example of a two-minute video with my mother cooking my
>> favorite dish, boeuf a la moutarde. She talks about Dijon but never says
>> the word mustard. Could I come across this video if you searched for
>> 'mustard'?
> 
> Nice example, but I don't buy it. If you're searching for this video
> with your mom in it, it'll have a variety of metadata with it. Dates,
> thumbnails, perhaps some tags or related videos. If "mustard" doesn't
> return it, a search for "mom", "dish" or "1998" might. Or just scroll
> through a bunch of videos with thumbnails and you'll probably find it.
> You don't need complete metadata, just good enough.

Where does "1998" come from? Does "mom" mention "1998" at any time?

Algorithmic extraction of metadata falls down when you really want to
capture the situation context, which can sometimes be very important and
useful metadata. It could be that the user remembers that there exists a
video clip (shot by uncle Tomm visiting from Sweden) of mom in the kitchen
preparing a dish (don't remember what) on the occasion of dad's 50th
birthday (not visible in the video). Being a home library of video there
could be hundreds upon hundreds of clips of "mom in the kitchen preparing a
dish" (it's her thing, she does it well), and since uncle Tomm is *behind*
the camera it's a tad difficult for an algorithm to figure that out. Sure
you could find it eventually using common clues and an exhaustive search ...
but the human brain has this habit of wanting to search by distinctive
things, such as that time uncle Tomm came all the way from Sweden to
celebrate his brother's 50th birthday.

I'm facing a similar problem in a project just now: we have masses of photos
and video of celebrities on red carpets ... and we need to have the database
record which event they were at (and which year), and which movie (if any)
they were there to represent, and possibly even which sub-event or other
related matter. Luckily we have most of this metadata already ... hand
generated by trained humans, not algorithms.

e.