Tuning of Comskip Part 2, ver 0.69

[erik]

From the information you mention in your post I do not have any suggestions except that I suspect the show_percentile.
Please post (or mail) the entire log file so I can have a look.

[erik]

During my work on using aspect ratio changes for finding cutpoints I have broken the scene change algorithm. (option 4)
Please do not use it for the moment.
It will work again in the next release.

[reboot]

I was wondering why there were logs of 5 meg or more, all filled with scene change stuff.
I don't want to clog up the forum with useless logs, that mean nothing to anyone but you and me, so I'll just keep testing for myself.
When I get 100% detection, maybe I'll write a tutorial on tuning (taming?) this beast, and how to get from the defaults, to a comskip.ini that will work.
Based on your suggestion about percentile, I am running more tests.
It seems that the detections are working extremely well, but when whole blocks are missed, using some deductive reasoning, I came up with the same answer...
I'm currently adjusting it, slightly higher, a bit at a time, and it's getting better.
I'm now at 0.73, because 0.71 isn't detecting whole blocks.
By law, there are only 12 minutes of commercials per hour. This does not include station identification (15 seconds per half hour), and parental warnings (60 seconds per hour). Total non-show time per hour is 13.5 minutes, so with no pre/post padding, that is 77.5%
According to the logs, the logo is displayed for ~74.x % of frames, but there are times when it's only 69.x%...it's difficult to narrow it down.
I also adjusted upwards, the max_brightness, and test_brightness, but left my max_avg_brightness very low, at 25.
This seems to take care of those random "sparkles" that show up, no matter what I do with my cables.

I do have another question though...it seems no matter what I set ar_split at, comskip detects my whole recording as widescreen, when it's obviously not.

If I haven't said it lately, I do thank you and Jere immensely for the ongoing help and support on this.
Maybe SA will take care of commercials, once and forever, but like a mad dog with a piece of meat, I'm not letting go, until I get comskip fixed to my satisfaction

[erik]

In the next version you can give the show percentile as a small number that then will be added to the logo percentage.
Thus if the logo is showed for 69% till 71% and there is always 4% of show without a logo you set show_percentile=0.04
Or you give show_percentile as a big number and then it will be used as is.

[reboot]

I'm having a terrible time with one show.
I have changed so many settings now, I'm lost.
I get near perfect detection on everything, except one commercial.
The actual commercial starts at frame 40589, yet comskip insists that it starts at 41177.

Black frames detected accurately at:

Code:

Frame  40579 - Black frame with brightness of 16 and uniform of 85
Frame  40580 - Black frame with brightness of 16 and uniform of 82
Frame  40582 - Black frame with brightness of 16 and uniform of 65
Frame  40583 - Black frame with brightness of 16 and uniform of 58
Frame  40584 - Black frame with brightness of 16 and uniform of 86
Frame  40585 - Black frame with brightness of 16 and uniform of 68
Frame  40586 - Black frame with brightness of 16 and uniform of 73
Frame  40588 - Black frame with brightness of 16 and uniform of 78
Frame  40589 - Black frame with brightness of 16 and uniform of 77

The next line is this:

Code:

End logo block 1    frame 40745    Length -  0:13:23.14

This is wrong. The commercial starts at frame 40589, yet comskip doesn't cut until frame 41177.
Other relevant portions of the log:

Code:

Block: 13    Start:  30693    End:  33437    AR: WIDESCREEN    AR_R: 1.80    Length:  0: 1:31.56
Block: 14    Start:  33438    End:  43291    AR: FULLSCREEN    AR_R: 1.58    Length:  0: 5:28.76
Block: 15    Start:  43292    End:  44194    AR: WIDESCREEN    AR_R: 1.96    Length:  0: 0:30.10

Code:

Joining AR blocks 16 and 17 because both have logo
Joining AR blocks 12 and 13 because both have logo

No mention of block 14 here.

Code:

AR Blocklist
Block: 5    Start:  13639    End:  43291    AR: FULLSCREEN    AR_R: 1.58    Length:  0:16:29.39

What's up with this? This block includes both show, and commercial.

Code:

Black Frame Count
110     40579     16        85
111     40580     16        82
112     40582     16        65
113     40583     16        58
114     40584     16        86
115     40585     16        68
116     40586     16        73
117     40588     16        78
118     40589     16        77

40589 is NOT a black frame, it's quite bright. the first frame of the commercial.

Code:

Creating block 27 From 40517 (40518) to 40579 (40578) because of Blackframe with 1 head and 1 tail
Creating block 28 From 40589 (40590) to 41176 (41176) because of Blackframe with 1 head and 0 tail

Something wrong here. Block 27 is correct. Block 28 is a commercial, yet detected as show.

Code:

Combining blocks 27 and 28 because they both have a logo.

This is wrong. Block 28 contains NO logo.

Code:

Logo start -  16675    end -  40745    length -  0:13:23.14    before:0.5 s     after:14.4 s

Wrong again. Logo start is 16675, but logo end should be ~40504.

Code:

The frames between  16661 and  41176 have an average AR of  1.00
Not trying to combine blocks 17 thru 25 due to excessive length - 535.201869
Not attempting to combine blocks 14 to 17 because block 14 has already been combined 12 times.
Block 17 Break for combined limit
Block 17 is shorter then minimum show segment.
Block 17 score:    Before - 1.00    After - 1.50

AR is correct, but again, the frame detection is wrong.

Code:

Block list after weighing
16)    4   2    0  16661  41176   0.00  817.985   0    0.00     0  1.58   4    FULLSCREEN                   Logo Present

Again wrong.

Code:

Final commercial cutlist
3 - start:  41177    end:  45402    length:  0: 2:20.97

Please help me sort this out somehow!

[reboot]

I also wanted to know how you got an AR of 1.45
PAL AR is 1.25:1 and NTSC AR is 1.5:1

[erik]

Looks like you have a false positive on the logo recognition.

With a well trained logo about 95% till 100% of the possible edges are actually seen each logo frame.
The default for a positive logo recognition is 75% or as it reads in the ini file
logo_threshold=0.75
Consider increasing this value a little, maybe to 0.80

Also remember that the logo recognition creates a block list ,separate from the aspect ratio block list, again separate from the black frame block list.
During the fuzzy calculation the three lists are combined.

Aspect ratio changes during logo are suppressed (that's why you see the "joining ar block" messages)

[King Mob]

I was having a similar problem where a commercial block immediately following the end of a show segment was registering as having a logo. Changing logo_threshold to 0.83 solved it.

Erik, it looks like you've added some tunable variables to the ini, as stuff like before_logo and after_logo are not on your list of variables you posted in the first Tuning Comskip thread. Can you list the variables and their defaults for v0.70? And what do before_/after_logo do when turned on?

[King Mob]

Also, wat command controls the modifier for the dictionary? I've changed the value for dictionary_modifier to no effect.

[erik]

I also wanted to know how you got an AR of 1.45
PAL AR is 1.25:1 and NTSC AR is 1.5:1

Thanks very much for this valuable comment :o
This explains why you in NTSC country do get WIDESCREEN instead of FULLSCREEN.

Setting the aspect ratio split value
The difference between WIDESCREEN and FULLSCREEN is determined by ar_split
ar_split=1.45
works for PAL because PAL FULLSCREEN is ar_ratio 1.25
for NTSC you could better use
ar_split=1.65
Look at the blocklist, if blocks are incorrectly tagged as WIDESCREEN you should raise the value.
If incorrect as FULLSCREEN you should lower the value.


Cutting before and after the logo
Setting before_logo to any value other then 0 will create a cutpoint (just like a blackframe) before the logo appears.
So
before_logo=10
will create a cutpoint 10 seconds before the logo appears unless there is a blackframe nearer to the logo.
The same is true for the after_logo
The minimum usefull values are 8 seconds or above because of the logo recognition time.
I have succelfully used 10 seconds for both, but that is VERY broadcaster dependend.