[mythtv-users] Patches that improve mythcommflag accuracy

[faginbagin]

Hoi Hika,

On 11/23/2014 6:28 PM, Hika van den Hoven wrote:
> Hoi Doug,
> 
> Sunday, November 23, 2014, 11:29:25 PM, you wrote:
> 
>> faginbagin wrote:
>>> Commercial flagging is a tricky business. It's relatively easy to tune it for a specific recording, but quite often what you do for one recording, seriously screws up the accuracy for another recording
> 
>> Too bad commercial flagging couldn't 'learn' from the end users.  For
>> the initial training, it would identify the show/series/what have you
>> and; we could go in and do corrections (training).
> 
>> Sort of like bayesian training for SPAM filters.
> 
>> Doug
> 
> I now have the patches for almost three weeks in place. Testing took
> some time for most I watch is commercial free. It looks like it gets
> all breaks, but all to short. So far I haven't noticed false catches.
> Most breaks are 7 to 8 minutes long and it catches only just over 6
> minutes. Since I have the skip set to manual I first thought that
> somehow the max length was set to short, but on checking it is as I
> thought set to 600 or 10 minutes. Skipping again doesn't work or skips
> to the next break.

There is a setting, CommDetectMaxCommBreakLength, that defaults to 395 seconds. It's possible that increasing it to 600 might help your situation. You would have to insert a record in the mythconverg.settings table to override the default.

> I don't know about the mechanisms involved, but most breaks start and
> end with the show sponsor clip, so I would think that would make it
> easier. Maybe this to short catching has to do with my problems with
> the seektables as mentioned further on, but I always rebuild them
> before checking.

In older versions of mythtv, I had noticed the seek table for some recordings would be missing entries for the last several seconds, sometimes minutes. I haven't been aware of that issue in 0.27, but maybe it's still an issue with low bit rate recordings? I'll try to find time to look into that.

> My settings are to use all methodes and no experimental speed-up or
> strickt checking. I record SD with two PVR500's on a dedicated backend
> with a AMD64 dualcore and 4Gb memory.

Please enable the strict checking option and then compare results. I have found accuracy is much better with it set, both without and with my patches.

> I's difficult to compare with the past for I have trouble since
> upgrading from 0.26 to 0.27 with my seektables (they make the show
> much to short) and I had an error in my script to first rebuild before
> comm checking. I only fixed it when I compiled with the fixes. So the
> last half year I skipped by hand ( I have right set to skip forward 1
> min. and left to skip back 10 sec. Very handy). In 0.26 I had regular
> falses and misses and also had sometimes (after I set it to manual
> skip) to skip twice. But that could be before I lengthened the max
> skip. But also often the full length.
> I live in the Netherlands so all shows are from the dutch commercial
> channels. If you want I can get you some sample recordings. Just say
> how to get them to you.

I appreciate the offer, but it's a time consuming and messy business to download other people's recordings and then insert reasonable records for those recordings in the database. So I would first prefer help in the form of feedback from people testing on their own systems. But if you are sure your results are consistently worse for most recordings with the patches than without them, then it might be worth the hassle to get a representative sample recording. 

Thanks,
Helen