Trim

To discuss the tuning of comskip and for posting your ini files
MCE_Anders
Posts: 12
Joined: Mon May 05, 2008 8:02 am

Trim

Post by MCE_Anders »

Hi

I have tried to trim Comskip, and I feel that the result is quite good, but... I dont understand how the cutlist is constructed since I have a perfect blocklist.

In this example I have a scoring as low as 0, and logo detected in the only blocks identified as show. Still Comskip includes more than that in the cutlist. I’d likes to prioritize SC and logo since they are very indicative. I tried to lower global_threshold but it messed up the scoring.

How do I achieve this?


Image

Image
erik
Site Admin
Posts: 3368
Joined: Sun Aug 21, 2005 3:49 pm

Re: Trim

Post by erik »

You are analyzing a 12 minutes recording.
Comskip can not produce reliable results on recordings shorter than 30 minutes as the global heuristics will fail.
Please try a longer recording and do post the log file generated with verbose=10 and indicate where comskip is making a mistake.
MCE_Anders
Posts: 12
Joined: Mon May 05, 2008 8:02 am

Re: Trim

Post by MCE_Anders »

Ok, I was following your "Tuning guide" and started to use short clips to tune Comskip. Maybe was a little to quick to jump to conclusions ;)

But as I can see, the conclusion matches reality quite well. For two channels Comskip always “allows” more commercials even though they don’t have logo, so I am interested in be more weight to logo detection. My experience says that I could almost use the condition: “Logo=show, No logo=commercial” and achieve very accurate cutting.

I have tested full length recordings, but I haven’t checked “the perfect cut” yet. I will do that in a few days.

The basic question still is: How can I weight logo and sc harder?

Another question: Using the GUI to see frames etc. (using p, n arrows) isn’t working very well. I can se the position, but the picture is all black so I can’t really tell what’s to view…

/Anders
erik
Site Admin
Posts: 3368
Joined: Sun Aug 21, 2005 3:49 pm

Re: Trim

Post by erik »

ComskipGUI should work on mpg and ts recordings but has difficulty on dvr-ms recordings.
How do you record?

When you want to give more impact to logo you should modify the "modifier" settings in the ini file.
They are documented in the tuning guide.
MCE_Anders
Posts: 12
Joined: Mon May 05, 2008 8:02 am

Re: Trim

Post by MCE_Anders »

I record as dvr-ms. Will convert my test recordings to mpg during the night to make it easier to proceed. Will be back!

I dont understand how modifing logo_present_modifier works. If I increase the value, I increase the scoring for blocks with logo. The blocks wíthout logo is not affected since the modifier multiplies with 0. As I can see I want to do the opposit: Increase the score for the blocks without logo. Have i misunderstood something?
erik
Site Admin
Posts: 3368
Joined: Sun Aug 21, 2005 3:49 pm

Re: Trim

Post by erik »

The modifiers do not work as add/subtract but as multipliers.
So any value above 1 move the score towards commercial and a value below 1 moves it towards show.
If you want the logo modifier to have more impact then divide the value by 10. Instead of 0.1 set it to 0.01

If you need some tuning help feel free to follow the instructions in the "how to ask for help" topic.
MCE_Anders
Posts: 12
Joined: Mon May 05, 2008 8:02 am

Re: Trim

Post by MCE_Anders »

IC

Thanks for your replies.
With 3 kids and a news hungry wife, there are not many minutes to tweak a Media PC ;)
I usually get some minutes by midnight...

I appreciate the offer, and maybe I’ll end up asking for help. For the moment I enjoy the challenge.
erik
Site Admin
Posts: 3368
Joined: Sun Aug 21, 2005 3:49 pm

Re: Trim

Post by erik »

I moved some recordings to another computer and did the comskip tuning there.
Once happy with the results I copied the tuned comskip.ini to the media PC.
WAF is VERY important. :D
erik
Site Admin
Posts: 3368
Joined: Sun Aug 21, 2005 3:49 pm

Re: Trim

Post by erik »

You asked for help and did send me a very nice set of test files.

Here is my proposal for improving the tuning

Parameters to consider when tuning this set of recordings.

Changed:
Disabling Heuristics 6 to avoid deleting the short commercial at the start
disable_heuristics=36

Changed:
Deleting part of previous or next show
delete_show_after_last_commercial=1
delete_show_before_first_commercial=0
delete_show_before_or_after_current=1

Not Changed:
Only delete a commercial at the start or the end if it is longer then this
min_commercial_break_at_start_or_end=39

Not Changed:
Delete blocks when there is no logo
punish_no_logo=1

Changed:
Filter the logo detection to eliminate short logo drops
logo_filter=5
MCE_Anders
Posts: 12
Joined: Mon May 05, 2008 8:02 am

Re: Trim

Post by MCE_Anders »

Great!

I was looking in my mailbox and my personal messages and it never occurred that you might answer in the forum :roll:

I really appreciate that you took your time to look into my setting. I will give it a try next week. This weekend I will go camping (without TV and computers!)
erik
Site Admin
Posts: 3368
Joined: Sun Aug 21, 2005 3:49 pm

Re: Trim

Post by erik »

Any results?
MCE_Anders
Posts: 12
Joined: Mon May 05, 2008 8:02 am

Re: Trim

Post by MCE_Anders »

Yes, there are improvements. But the "spot on" expression that some users her write about, doesn’t occur for two channels (Basically most of my recordings are from 4 channels).

TV3 and kanal 5 behaves differently than TV4 and TV6. TV4 uses black frame before and after commercial brake, but the other two often have a channel specific scene with sponsor commercial and then a black frame.

I was wondering about "Disabling Heuristics 6 to avoid deleting the short commercial at the start
disable_heuristics=36"...
Which commercial did I want to keep, and should the value be 36 or 6?

Is there some documentation what those heuristic rules does?

I’ll keep playing with my recordings and will get back to you if I have more questions.

I still see it as a exciting challenge so I don’t mind to struggle for a while ;)

BTW: I used to have a automated process to cut the commercials out, so unfortunately I don’t have any of my favourite shows "uncut” left and now, most of them has ended for the season. As I mentioned earlier I find that detection differs slightly from prime time shows to afternoon shows so I would rather test on my favs. Now I changed the process so I leave the show uncut so I can use them for testing.
MCE_Anders
Posts: 12
Joined: Mon May 05, 2008 8:02 am

Re: Trim

Post by MCE_Anders »

Got some time and made a new set of files:

http://www.femtiosju.se/MCE_Anders.zip

There is one file called notes.txt that describes all the deviations in one single file.

First of all: I’m quite satisfied with the results, so there is no need for a lot of effort. I am mostly curious how perfect result one can get. Therefore there is no rush in answering either.

Some things that I noticed:

Even though Comskip finds frames from the beginning (like 1-1456) the XML-file that is produced never marks the initial commercials…

When they wrap the show in a “Show-look-a-like commercial” there are some trouble.

The cutpoint for TV3 is almost always 2 frames too late.

When there are previews for other shows, Comskip mark them as show, how strange ;) Is there a way to avoid that.

Otherwise I hope you could give me some clues to further tuning. I am most impressed by your skills and find it amazing to just skip commercials and concentrate on the shows. One gets really spoiled!

Regards

Anders from Sweden

EDIT
Tried to run the .csv file and got a .dif file. Very interesting! But I cant find out the codes that indicates the reason to the difference:
EL v b
EL b
EL s
/EDIT
erik
Site Admin
Posts: 3368
Joined: Sun Aug 21, 2005 3:49 pm

Re: Trim

Post by erik »

Nice set of test files.
Some improvements are still possible. Will give it a try
MCE_Anders
Posts: 12
Joined: Mon May 05, 2008 8:02 am

Re: Trim

Post by MCE_Anders »

Hi again!

Browsed the forum and found the answer for the missing part in the XML-file (upgrade to 79.132 i.e replaced the exe-files). When I did that I made a rerun for some files and surprisingly the results differ quite a lot sometimes. Unfortunately the results are less accurate. The Frusna_Kanal5 show detection was way of.

A complete new set of files is uploaded and in the file Differences_112_132.xls I have pasted the different cutpoints side by side from the two versions:

http://www.femtiosju.se/MCE_Anders2.zip

Since I’ve missed the correct format of the XML-file it sees to be a good idea to tune for version 79.132, but in that case it's a step back I’m afraid…

/Anders
Post Reply