[mythtv] How you can help Commercial Flagging development (was Commercial Flagging idea)

[Lucas Meijer]

Well, its hard to make serious contributions to the commercial flagger 
without being able to program, but there are a few things that would help:


- if the commflagger misflags a recording, find out why. you can get a 
lot of verbose information from mythcommflag if you ask for it. Does 
logo detection fail, does your blank frame detection fail, do you have 
no, or zillions of scene changes? Look at the framemap thats being 
output at the very end for an overview of the analysis the classic 
commdetector performed. Knowing where the current algorithm brakes down 
helps in fixing it.


- if your math is good, talk me trough the math at section 3.3 of
http://www.cs.cityu.edu.hk/~cwngo/paper/thesis-ngo.pdf

I've looked at all the academic research (and non academic for that 
matter) I could find, and this method seems by far the most 
sophisticated for scene change detection. I'm working on a test 
implmentation of this algorithm, but am stuck very early in the process 
:), not being fermiliar with complex gabor images. getting some 
assistence in this area would be great.

- if the commflagger misflags a recording, keep that recording aside or 
make a backup for it, so when new developments kick in, we get a 
testbase we can test against.

- fresh ideas are always welcome, alltough a lot of them have already 
passed the list and are not picked up (yet) because noone that can code 
wanted it badly enough. so search the archive first.


As for current state of things, Chris Pinkham maintains the Classic (and 
only) commercial flagger. I've been doing work that nobody has seen yet, 
since it breaks down due to poor current scene change accuracy 
(especially fadein's & fadeouts get marked as a huge amount of scene 
changes, corrupting useful statistics).

I am hoping to adapt the 'plugin' architecture so that a 
commercialdetector plugin can also access the audio data.

Another ideas I'm working on but didn't materialize yet is trying to 
find those 2-4 second tv station "videojingles" they put in front of 
commercials, by looking for scenes with identical durations, and then 
checking their histogram/image, and maybe even their audio..

Bye, Lucas