[vdr] Re: Proposal: cutting out advertising automatically

[Margit.Fiegert@t-online.de (M. Fiegert)]

Harald Milz wrote:

>Hi,
>
>now that we have an MPEG-2 stream on disk, has anybody thought about
>(semi-) automatically deleting all the advertising blocks from the
>recordings?
>
>Some ways of doing the job come to my mind:
>
>  
>
...

>2. We could analyze the MPEG-2 data stream for a number of sensible
>criteria (sender logo present; average loudness; ratio of scene change;
>etc, anything else?, then produce a probability for commercials) and
>produce a suggestion for cutting. 
>
>3. We could set up a data base to collect cutting data. Everybody cutting
>a video could upload the data to this server, and everybody else could use
>it. Something like freedb.org or so. Like P2P networks this may result in
>some people uploading fake data to disrupt the service, though, so that
>proper encryption schemes may be needed. Even then, somebody may subscribe
>and upload fake data. The system may need a voting scheme in addition.
>
>At the moment, #2 seems to most feasible and least disruptible way to go.
>Whether or not and how the film industry reacts is a different question -
>but in Europe there seems no way of legally jeopardizing this method on
>open source systems. It also is no circumvention device (a la DMCA) because
>it does not illegally provide access to something.
>
>Any thoughts?
>
>  
>

Lot`s of thoughts, but again - lack of time for implementation up till now.
Just a glimps of one of my favourite follows.

Btw:
I agree with you Harald, that 2 is most promising.
I don`t agree with Mathias who wrote about 3: "-> Forget it. Not worth 
thinking about it. "for different reasons.
(two of them: we have the same digital streams, so a short sequence is 
enough to synchonise; my recordings usually are stored for weeks before 
I can watch them.)

My favourite would be a small fingerprint right out of the whole picture 
information (maybe directly out of mpg) every n-th I-Frame.
One such hash value is surely not enough to characterise an add, but the 
sequence is.

If you define advertising as "short (~10-50sec) sequences comming over 
and over again and always in groups."
then you`d need only two starting points (for example common 
advertisements or the typical  jingle before/after) to pull yourself out 
of the swamp.
If you find those, they are ads. If the gap in between is say less than 
10min, these are also ads.
By this a database of ad-fingerprints can be build up automatically (and 
not only while recording, but also in idle time)(and shared!).

To be safe in the beginning all scenes rated as adds could be stored in 
a backup dir, and only deleted as soon as probability becomes really 
high (eg. same 30 sec seen on another network in another transmission 
surrounded by different stuff). Then, in case of emergency one has a 
chance to get back the last 10 minutes of a thriller the quest would 
follow you the rest of your life otherwise ;-)

For a job I'm in corporate research.  Analysing and fingerprinting 
(usually noisy sensor-) data is in my core buisness.
So if anybody can offer time, I`ll be very happy to join the team!

Michael







-- 
Info:
To unsubscribe send a mail to listar@linuxtv.org with "unsubscribe vdr" as subject.