Comskip 0.79 build 24: Improved logo recognition

[erik]

I have posted a new Comskip version

It can be downloaded from here


From the readme.txt
Changes in 0.79.24
- Made a big change to improve finding weak logo's (like on NZ's channel ONE)
- Some heuristics improvements for very short segments
- Comskip will check your logo_threshold setting and suggest a value in the log file. reliable on recordings of at least 45 minutes.
- In case of a Comskip crash there is no longer need for manual intervention.
- Fixed a bug that could cause the wrong ini file to be used

[Lindsay]

Thanks once more for your on-going efforts, Erik. Your ini file is very close to what I run with now. You recommended I change logo_threshhold down to 0.60 because of the weak NZ channel ONE logo. Since you have done some work on this in this version, should I change the threshhold back up to, say, 0.80?

BTW, I am getting excellent results on my main recording channel (ONE). I am a very pleased comskip user.

[erik]

Don't change the setting, Comskip calculated the same value for you recording

[Lindsay]

Ah, I see!!

Thanks, Erik.

[erik]

Tipstir

I am happy to see it works works for you.
Setting max_brightness=255 effectively disables some of the black frame detection algorithm.
The main reason why someone may do that is because part of the logo is not inside the logo box or there is some other permanent bright part on the screen.
I would not recommend this settig in general.

[thawk]

Thanks for this version - I have got it working pretty well now - I tried Showanalyzer but couldnt get as good results for the UK, partly I think because I could find less variables to play with and partly perhaps because the testing environment for Comskip is now v effective.

I am running GBPVR in the UK with input from an NTL cable TV box via a PVR150. Using the postprocessing file I am skipping processing channels with no ads (BBC), not using any logo processing on channels with no logos (nologo) and otherwise using a very logo-focused approach for all the rest with logos (the only exeption being More4, a channel which has a very faint logo that even on the most sensitive settings is only sometimes picked up)

There is however an error I think in the postprocessing bat file that comes with the Comskip zip file - there shouldnt be so many colons around the labels I think! eg mine now reads as below and works:

if "%2" == "1" goto eof
if "%2" == "2" goto eof
if "%2" == "10" goto eof
if "%2" == "12" goto eof

if "%2" == "3" goto nologoch
if "%2" == "4" goto nologoch
if "%2" == "5" goto nologoch
if "%2" == "9" goto more4

goto generic

:nologoch
comskip --ini=nologoch.ini %1
goto continue:

:more4
comskip --ini=more4.ini %1
goto continue

:generic
comskip %1

:continue

:eof


In terms of the settings files, for the channels with no logos (terrestrial channels ITV, C4 and C5) which have ads of around 4mins pretty much on the hour, 1/4 of hour, 1/2 hr and 3/4hr, I use the following ini file - it works around 90% of the time except sometimes near the start or end of programs - I have to use scene change rate to pic up the ads (otherwise there is little except timing to distinguish from show)

However sometimes this scores highly a bit near the end or start of the show and then links it in to the ads including any bits with lower scores too that definitely are show in between. Is there any way of making sure sections with scores above the threshold are not linked together near the start or end of the file?

Also I get the impression in the scoring output files that sometimes blocks are not scored for length, just for scene rate change - does the scene rate change scoring override the length scoring? For these channels and blocks I would expect scoring by both for all blocks would be more accurate?

Anyway this is the ini file for these channels: nologo.ini

detect_method=45 ;1=black frame, 2=logo, 4=scene change, 8=fuzzy logic, 16=closed captions, 32=aspect ration, 255=all
verbose=0
max_brightness=60
test_brightness=40
max_avg_brightness=25
max_commercialbreak=300
min_commercialbreak=120
max_commercial_size=65
min_commercial_size=6
min_show_segment_length=600
non_uniformity=1000
max_volume=100 ; any frame with sound volume larger than this will not be regarded as black frame
ticker_tape=0
intelligent_brightness=0
global_threshold=1.2
padding=4 ; ensures 4s added to start and end of show segment so that more chance of seeing start/end of ad to check all ok
delete_logo_file=0 ; set to 1 if you want comskip to tidy up after finishing
output_framearray=0 ; create a big excel file for detailed analysis, set to 0 to disable
output_videoredo=0
output_debugwindow=0
output_womble=0
output_cuttermaran=0
output_mpgtx=0
output_zoomplayer_chapter=0
output_zoomplayer_cutlist=0
output_edl=0
output_vcf=0
live_tv=0

For the majority of other channels on cable in the UK, the logo is clear and present only during the show - and always present during the show! Hence the default ini file is heavily weighted towards logo detection and scoring - this works basically 100%. The key things I played with were for increased sensitivity of logo detection (most logos are white watermark variety - there are almost no '0800' bits on ads or other things to confuse detection):

logo_threshold=0.70
logo_max_percentage_of_screen=0.1
edge_radius=8
edge_threshold=2
give_up_logo_search=600

I also deliberately played with settings to increase the number of 'black' frames registered and break points:
max_volume=200
non_uniformity=1000
cut_on_ar_change=1 ;also insert cuts on aspect ratio changes
brightness_jump=80 ;this I believe introduces cuts when there is a reasonable brightness jump between frames - default is 255 I understand

The reason for this is that on many of the logo channels shifts from ads to show and visa versa occur sometimes with no black frame (!) so I seemed to need to increase the likelihood of introducing a break that comskip could grab hold of. I then made sure the larger number of fragmented blocks with logos were joined and scored very low (ie as show) and the rest as adds:

connect_blocks_with_logo=1
logo_present_modifier=0.05

Overall works really well - this default comskip.ini is below:

detect_method=47 ;1=black frame, 2=logo, 4=scene change, 8=fuzzy logic, 16=closed captions, 32=aspect ration, 255=all
verbose=0
max_brightness=60
test_brightness=40
max_avg_brightness=25
max_commercialbreak=400
min_commercialbreak=120
max_commercial_size=120
min_commercial_size=19
min_show_segment_length=280
non_uniformity=1000
max_volume=200 ; any frame with sound volume larger than this will not be regarded as black frame
ticker_tape=0
intelligent_brightness=0
logo_percentile=0.92 ; if more then this amount of logo is found then logo detection will be disabled
logo_threshold=0.70
logo_max_percentage_of_screen=0.1
edge_radius=8
edge_threshold=2
aggressive_logo_rejection=0
brightness_jump=80
give_up_logo_search=600
connect_blocks_with_logo=1 ; set to 1 if you want successive blocks with logo on the transition to be regarded as connected, set to 0 to disable
cut_on_ar_change=1 ; set to 1 if you want to cut also on aspect ratio changes, set to 0 to disable
delete_show_after_last_commercial=0 ; set to 1 if you want to delete the last block if its a show and after a commercial
delete_show_before_or_after_current=0 ; set to 1 if you want to delete the previous and the next show in the recording, this can lead to the deletion of trailers of next show
shrink_logo=0 ; Reduce the duration of the logo with this amount of seconds
after_logo=1
before_logo=1 ; set to number of seconds after logo disappears comskip should start to search for silence to insert an additional cutpoint
padding=1
logo_present_modifier=0.05
delete_logo_file=1 ; set to 1 if you want comskip to tidy up after finishing
output_framearray=0 ; create a big excel file for detailed analysis, set to 0 to disable
output_videoredo=0
output_debugwindow=0
output_womble=0
output_cuttermaran=0
output_mpgtx=0
output_zoomplayer_chapter=0
output_zoomplayer_cutlist=0
output_edl=0
output_vcf=0
live_tv=0

The real problem I have is with More4 channel which has a very faint inverted watermark logo (ie the logo pattern is darker than the picture!). For this one I have tried to further increase the logo detection sensitivity (assuming that shifting the numbers in these directions makes it more sensitive which it seems to - with these it sometimes but still v rarely picks up other things on the screen than a logo, but still sometimes misses the logo especially on things like West Wing which is very dark anyway!):

edge_radius=12
edge_threshold=0

So overall Comskip can work well, especially on logo channels in UK (I have seen coments suggesting it can't!)

However:
- Any ideas how to better pick up the inverted dark logo in More4?
- Any ideas how to better resolve issues at the start/end of programs with no logos - is the block scoring system for length always working vs for scene change?

Thanks again for a great program!

Tim

[erik]

I am very much impressed by this feedback!

W.r.t. to faint logo's
If you can learn the logo on a specific show then you can reused that learned nogo on other shows on the same channel when the logo stays at the same position. This does not have to pixel accurate but it will not survive a 4/2 to 16/9 transition.
Use the --logo= parameters to specify the logo file so
comskip --ini=inifileforthischannel.ini --logo=logofileforthischannel.logo.txt
This prevents comskip from not finding the logo.
I am working on a automatic setting of the logo_threshold, this may also help on faint logo's

W.r.t. no dark frame's
I am working on another detection strategy that does not require dark frames at all. First results vary over channels but on some they are very good.

W.r.t. scoring on scene change
The scoring on length is tuned to certain advertisement styles. UK may have others (I have seen one USA channel with 36 seconds advertisements) Could you send me one or more of the .csv's with a corrected .txt of show's that is causing problems? I like to have a look to see what is possible.

W.r.t. colons
You are right about too many colons

I never tried these settings.
edge_radius=12
edge_threshold=0
Specially the last setting does not seem to make sense.
Could you upload about 400MByte with the More4 logo to my ftp server?

[thawk]

Thanks for reply

I have deleted some of the examples that gave probs but will send the .csv and upload the next time I get one

I have an episode of West Wing from More4 which I can upload where it found it impossible to find the logo! I will try your idea of sticking a logo file that it did find in.

Where/how do I upload to your ftp server?

Cheers

Tim

[PVR4ME]

I have been reluctant to install any type of commercial detection s/w for two reasons:
1.) Commercials can be manually skipped fairly easily during playback
2.) The documentation for comskip, at least, looked pretty complicated and left many questions for the "newbie" like me.

I must say, though, that I took the leap to install and play with comskip and was amazed at how well it worked right "out- of-the-box" so to speak. The integration w/ GBPVR is excellent and the default settings worked great for me. I've only ran it on a random show I picked just to see what it would do and I am very impressed with the results - great job guys! I really like the option in GBPVR to auto skip commercials (without having to cut and paste a new file) and it's cool that the timeline reflects the commercial gaps. Good stuff! Thanks to all involved!

[firehorse]

Hi Tim,

Thanks for reply

I have deleted some of the examples that gave probs but will send the .csv and upload the next time I get one

I have an episode of West Wing from More4 which I can upload where it found it impossible to find the logo! I will try your idea of sticking a logo file that it did find in.

Where/how do I upload to your ftp server?

Cheers

Tim

I was wondering if you had made any further progress with your comskip settings

Thanks
Alan