Detect_method education

[vancedailey]

Hello,
I am looking for some help in selecting the optimal set of detect_methods. I realize there is no "one size fits all" or else the choice would not be left to the user. However, I hope I can get some generalizations that can guide my experimentation.

1. I only record HD shows. Are some methods unnecessary or less effective on clean HD recordings?
1a. Can some sensitivity settings be tightened up for clean HD recordings?

2. Are some methods more prone to mis-identify a portion of a show as a commercial? (very bad)

3. Are some methods more prone to mis-identify a commercial as a portion of a show? (less bad)

4. Do certain combinations of methods work well together because one method succeeds in cases where another method fails?

5. Does the addition of more methods increase the likelihood that more commercials will be found (and increase the risk that portions of the show will be mis-identified as commercials?

To clarify the last question, does fewer detect_methods=fewer commercials found and more detect_methods=more commercials found?

I also have a few specific questions about each method
1=black frame
I assume the there are a few black frames before and after each commercial.

2=logo
The rule I guess is that commercials don't have a logo while shows might.

4=scene change
An abrupt change from one frame to the next might signal the start/end of a commercial.
It would seem that scene changes by themselves would not be very predictive.

8=fuzzy logic
My guess is that this is not a true detection method.

16=closed captions
Is it the change in a caption that signals a possible transition from show to commercial?
Does this work if I have captions turned off? (is it still recorded but not visible?)

32=aspect ratio
Is it the case that commercials are sometimes displayed in 4/3 even with the show is in 16/9?
This method would seem to work less well on the major network HD channels.

64=silence
Does this work better with HD audio formats?
Does this work well with 1=black frame?

128=cutscenes
How does this method work?
Is it related to 4=scene change?

Thanks,
Vance

[erik]

Please read the available documentation and search the forum as you ask too many answers to be anwered in one post

[vancedailey]

I realized it was a stretch to ask so many questions at once

I read your tuning guide and a few others I found on the Internet. The information is helpful but your program is still very much a black box to me. I have 4 tuners and 1000+ channels so the thought of trying to tune individual channels or shows is daunting. And if I tweak one show/channel I imagine I risk fixing the detection on some recordings and breaking it on others.

I hoped that figuring out my optimal device_method was a good place to start.
So, I figured I would fish for some guidance on how to select a set of device_methods.

Can you point me to a guide or thread that discusses device_methods?

Thanks,
Vance

[erik]

Its much more complex than setting the detect method.
If you want something that works on 1000 channels you need to get representative recordings from all important channels and optimize many parameters, or you simply use the default ini file and only try two changes

detect_method=111

and

min_show_segment_length=250

Succes will depend on the mix of broadcasts your find most important.

[vancedailey]

Faced with choosing between a 1 minute change or hours of experimentation I choose trying your suggested settings.

Five years ago I considered using Comskip but never followed through.
I did however pitch an idea for another detect_method which failed to excite your interest.

Let me pitch it again.

The idea was that since most commercials are shown many times you might suspect that a 30 or 60 second segment of video was a commercial if it was a duplicate of another segment of video. If you could generate and store some sort of a video checksum to recognize two nearly identical video segments you would have another clue that you were looking at a commercial.

I doubt that its an option but if there was a way for everyone to share their checksums...

Thanks,
Vance

[erik]

That reminds me of the bowflex commercial........
Its an interresting idea but at the moment I have other high priority features to implement.

[bmblank]

How would that detect method deal with reruns? If somebody had had watched a show, then shared the checksum, would that not call my whole video a commercial, were I to watch the same thing?
It would be neat. And if it worked right, seems like it'd be bulletproof.

[erik]

I know how to do that

[bmblank]

I'm sure it was evident already, but if not, now its painfully obvious that you're way smarter than I am.

[vancedailey]

Your reply to my idea back in 2007 was:

"Having this kind of finger printing is possible but is a project in itself.
I rather not spend any time on it for the moment.
There is little time and many other usefull things to do."

I just have to catch you in a slow period.
I'll check back in 2017

Keep up the good work.
Thanks,
Vance

[vancedailey]

Setting the detect_method to 111 has helped. Thanks for the suggestion.

I use Recorded HD TV (http://www.recordedtvhd.com/) and it shows the sections of the show marked by Comskip as commercials. This allows me to visually check a recording and see if the detection looks reasonable.

Frequently, I can tell just by looking whether Comskip guessed correctly.
If a human can spot patterns maybe Comskip can too.

Commercials seem to generally be evenly distributed throughout a show. (Exceptions abound of course)
Do you award extra points if the commercials seem to be spaced 10 minutes apart and the segment you are evaluating is 10 minutes after the previous segment identified as a commercial?

You have minimum and maximum values for show segments and commercials but do you use a sliding scale to determine the probability that a segment is a commercial or show. For example, you may define 250 seconds as a minimum show interval but a 300 second segment is more likely to be part of the show than a 251 second segment. And a 350 second segment is even more likely to be part of the show.

Thanks,
Vance

[erik]

When you set
verbose=10
you can see in the comskip log file how the whole process of detection works.