[mythtv] A thought on Commercial Flagging -- Long (And rambling)

Robert Johnston anaerin at gmail.com
Sun Aug 7 16:39:04 EDT 2005


On 07/08/05, Lucas Meijer <lucas at mach8.nl> wrote:
> Every now and then the "lets make a big online database of commercials"
> albeit learning or not learning topic comes up on the mailing list.

The "Online Database" wasn't really the point of my message/rambling.
It's more of a "Once it's working, we _could_ do this" not a "This is
essential".

> It might not be a completely bad idea per se, but it's just way out
> there at the moment. Commercial flagging would be way better of with
> someone spending some time on better "feature detection"..  Chris has
> written pretty much everything himself, and while it's remarkable that
> things are often working, a lot of the detection (logo detection, scene
> change detection, blank frame detection) is not shockingly accurate.

I'm not meaning to demean what Chris has done. Most of the time, it's
accurate and works well. But there are times when it doesn't work, and
when it breaks it REALLY breaks. 47-minute commercial breaks are...
"Unlikely", yet the flagger seems to detect them often on Bravo,
mainly because they don't use an on-screen logo.

What I'm suggesting is that rather than having to keep adding bits to
the commercial flagging routines every time the broadcasters change
their tack, which ties up development time that could be used on other
things, we have a "Learning" system that will adapt "on the box" to
improve commercial detection in a continuing and ongoing way, without
having to rewrite, rebuild, or recompile.

> If you want commercial flagging to get better, improve the methods that
> the highlevel algorithm bases its decisions on.

This is what I'm talking about. I'm suggesting using a "Learning"
algorithm for commercial flagging to work from, which will learn and
constantly improve itself based on user's input, tailoring itself to
their viewing preferences and wants.

> Anyone who can code can imho better spend his/her time on that, than on
> some self learning, global online commercial identification database. I
> am convinced "simple" look-at-video-and-audio commercial detection can
> become incredibly accurate if only more people work on it. Right now all
> we have is pretty much the work of one guy.

I agree that we need more eyes on this area, more people helping out.
I'm just offering a suggestion as to how we can "Solve" the commercial
flagging problem once and for all. (And offering some other ideas as
to how we can improve other parts of MythTV while we're at it).
-- 
Robert "Anaerin" Johnston


More information about the mythtv-dev mailing list