Re: TREC-10 Video Track - 2 updates, 1 question

["R. Manmatha" <manmatha@cs.umass.edu>]

Hi Paul and the Videogang,
	My two cents:

I suggest that we look at tasks we would like to do and worry later
about whether  we will be able to do these with the 
videos available. Also since the goal is to evaluate what
people are doing, we probably want to focus on what people
can currently do. 

Here is an example of possible tasks:

  Detecting duplicate or near duplicate video clips.
For example, assume that there are multiple examples of
the same video clip which differ in small ways. Some of
these occur naturally - like a shorter version of an
advertisement or a longer version of a news item. They
could also differ in other ways like missing frames or
the order of the shots could be scrambled slightly. The 
task will be given a query to find other duplicate or 
near duplicate clips.

It will be easy to create a database - just scramble existing
video clips. Also it would be nice to have a large 
collection of video clips for this purpose. This would
be a natural video retrieval task.

The same task could have easy and difficult versions.
For example, the easy version would involve finding video
clips which are much closer to each other and very
different from other video clips and so shot based or frame based
methods would work. 

Harder versions could involve selecting from one of a set
of video-conferencing clips - e.g. find me all the 
video-conferencing clips in which Richard Feynman is giving
a talk. In this case, face detection and face recognition
might be required or we may be able to exploit text.

For the easy versions we might require that other media like
speech etc not be used - it would be a test of things like
shot detection and similarity matching in essence.

The more difficult versions may allow other media.
I guess the more difficult versions are more like similarity
detection than duplicate detection.

There are more difficult video retrieval tasks that one might
want to do but it is unrealistic to expect to be able to do
them right now. For example, helping a film maker retrieve
shots of tigers for their documentary is probably quite
difficult to do at this time (assuming its not completely annotated with
text).

Comments?

Regards
Manmatha

P.S. Yes, I think it would be useful to have MPEG-2.