Re: master shot ref

[Georges Quenot <Georges.Quenot@imag.fr>]

Paul Over wrote:
> 
> Hi Georges,
> 
> I think there should be a minimum threshold on the size of
> shots we accept as search and feature extraction results
> because shots shorter than 3 secs or so cause problems for
> our assessment software, which uses a Real version of
> the videos. Also because some features require a minimum
> length due to their definition or to be perceivable.

I have some alternative solutions to Real to display shots
that can do with shots shorter than 3 secs. One is based on
the mpgtx (http://mpgtx.sourceforge.net/).

> Let's postpone setting that limit for the moment and
> separate this threshold from one that might be used in
> remaking the master shot ref - even though the two are
> related.
> 
> It's hard to know what to do about the master shot ref.
> I think we could live with the current one if we have
> to but it's worth thinking about a revision.

I have started a new segmentation, it should complete during
the night. The recovered transitions are identical to the ones
of the previous segmentation but they now come with a confidence
measure which allows then to discard the weakests of them that
do not fulfill the minimum spacing requirement. I have also done
the software for that and I tested it on last year SBD data. It
seems to work well. The remaining transitions are a subset of the
previously selected ones and the minimum shot length can be set
to any number of frames.

> As you point out, we could use the one we have. Then some
> small percentage of frames are lost to us because they
> belong to shots too short to even be used. More worrisome
> is the fact that the false short shots may chop real
> longer shots into pieces we cannot use. But I have no idea
> how often that is happening.
> 
> The best thing would be a rerun with the sensitivity knob
> turned down.  But maybe such a knob does not exist on the
> CLIPS-IMAG system or its not clear where to set it - thats
> a research issue for SB detection systems. So you must do
> some merging of weak transitions.

The problem while lowering the sensitivity of the systme (this
is possible) is that this removes many good transitions while
not necessarily removing the ones involved in short shots
oversegmentation. Therefore, leaving the system work as it is
and then remove the weakest transitions between short shots
is probably better. It seem also that there are some actual
shots that are very (short < 10 frames), these will be merged
with adjacent ones.

> If you do this, I think a larger minimum threshold of say 1 sec
> should be used. This makes the reference less of a true shot
> reference and the accuracty of the evaluation a bit less. But
> the shot reference is mainly a means to common evaluation based
> on some logical segmentation and that goal is not violated if
> if two shots are merged now and then. The granularity of the
> evaluation is coarser but still useful I think.

I will generate two new segmentations with a minimum shot
duration of 15 and 30 frames respectively. You can then choose
the one you think is the more appropriate. Both should be ready
tomorrow morning. I will send them to you and Jiamin Ye so that
he can rebuild the MPEG-7 compliant version.

Once the segmentation with confidence measures is finished,
I can build easily and quickly a segmentation with any minimum
duration.

> Even if there are no objections to this - I'm thinking
> especially of the groups working on ASR results for
> July 1 and tied to the master shot ref - I am not clear
> about how you will do the merging, either the simple
> or more complicated way. Can you explain?

I finally did an intermediate choice. I do the segmentation
once with quite low thresholds (as it was originally) while
associating a confidence measure to extracted transitions.
Then I remove the transitions surrounding the strongest ones
and located at a distance less which is than the minimum value.
A simpler way would have been to simply remove the transitions
located too close to the previous one regardless of their
confidence. Many more complicated methods can be thought of
but does not seem necessary.

> PS: I think the percentages you gave for the 2001 SBD
> collection shots should be 10 times larger - but the
> point you make with them is unchanged.

This is correct. The counts were related to the total number
of transitions of the feature test collection (~35000) and not
to the total number of frames of last year SBD collection
(~3000). The parameter was not updated.

Georges.

Email: Georges.Quenot@imag.fr
CLIPS-IMAG, 385, rue de la Bibliothèque, B.P. 53, 38041 Grenoble Cedex 9
Tel: (33-4) 76 63 58 55, Fax: (33-4) 76 44 66 75