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

Robert Johnston anaerin at gmail.com
Sat Aug 6 01:52:04 EDT 2005


I'm sure this has been suggested before, but if it hasn't, here's an
idea I came up with while I was trying to get to sleep (It's far too
hot for that, really, but what the heck).

Anyways, How about we take the principals of Neural Networks or
Bayesian Filtering and apply it to Myth Commercial Flagging.

Now, I know Bayesian filtering's been suggested before, albeit for
programme selection, but hear me out on this one.

If we (by which I mean "You", as there's no way I have enough
programming experience to really start with C++, let alone Neural
Networks) could develop a system that "Learns" what Ads are, we'd have
a good start.

Then, to "Train" the system in an ongoing manner, we add hooks into
the "Playback Recordings" system.

So, for example, if someone is watching a show, with Auto-Skip set on,
MythTV skips a commercial, but also skips part of the show, the user
will rewind from that commercial point to see what they missed. We
could take note of that and use that information to say "Detection was
accurate up to x, but we misidentified y as an ad".

Similarly, if detection comes up short, we can detect that the user
fast-forwarded from the "End" commercial point and use that
information to say "X was good, but y is also an ad".

We can also have a reasonable "threshold" for these detections. For
example, if the user does nothing for more than five seconds after an
ad-skip (If it's manually triggered. Perhaps ten seconds for an auto
skip to give them time to find the remote), we can assume that all has
gone well, as the user isn't having to correct the Commercial
Flagger's mistake.

We can then (optionally) take these "Learned Rules" and send them
(anonymously) back to a central database, so every myth box
contributes and the whole Myth community acts as a very big (and very
intelligent) Distributed Neural Net.

Yes, I was thinking all this up at 11:46pm, instead of sleeping. My
name is Robert Johnston, and I am a geek.

Now, this could all be a massive, wild stab in the dark, or I could be
off my rocker, or any number of other things, but this is only an
idea, nothing more.

For the very initial training, I was thinking of a system where you
have a keyboard button set up to mean "This is an ad", and sit
watching hours of "Live TV" with the box, holding it's hand through
several hours of programming, to teach it what Ads are. I realise that
this would be very tedious and menial work, but it should only have to
be done once, by the developer of the system, and his "Base Data" can
be used as a starting-point for ad-detection on all other systems.
Something to train from.

As I said, just a thought.

Comments? Suggestions? Flames?
-- 
Robert "Anaerin" Johnston


More information about the mythtv-dev mailing list