Given how important logo detection is to comskip, it seems like having a good logo.txt file for your favorite channels would speed up and improve comskip. When I look through a series of logo.txt files from a given channel, they all vary slightly. How can I tell which one is most accurate? Also, would it be worthwhile to establish a reference set of these for use by U.S. comskip users?
Thanks!
Logo files
When you look in the .log file that was generate while creating the logo file you will see the amount of logo found per block at the end of the log file.
Look at this example under the word "logo"
The closer to 1 the value is under "logo" where there is supposed to be logo the better. For the blocks where there is supposed to be no logo the value onder "logo" should be 0
I am willing to host a library of logo files. Let me know what kind of information you would like to record with the logo file. I was thinking about the resolution used when recording, the name of the broadcaster, channel number? Amount of border used? Aspect ratio of broadcast ( the logo position often depends on this aspect ratio)
Look at this example under the word "logo"
Code: Select all
Block list after weighing
----------------------------------------------------
# sbf bs be fs fe sc len scr cmb ar cut bri logo vol sil corr stdev cc
0:-- 10 0 10 1 390 2.00 12.980 0.29 0 1.60 L b 2089= 0.00 1751+ 89+ 0.000 208
9:-- 10 8 2 4812 5706 9.00 29.830 0.91 1 1.60 LC S b 6498= 0.00 1881= 98+ 0.000 649
10:++ 4 3 1 5707 10669 0.11 165.566 0.08 0 1.60 EL b 798= 0.84 1878! 95+ 0.000 79
11:-- 9 1 8 10670 11576 7.88 30.230 0.36 1 1.60 LC S b 2941= 0.00 1981= 97+ 0.000 294
12:-- 11 8 3 11577 11731 3.94 5.138 0.01 1 1.60 LCN a b 58! 0.00 1057! 57! 0.000 5
13:++ 3 3 12 11732 33990 0.07 742.676 0.10 0 1.60 EL b 886- 0.92 1546- 850 0.000 88
14:-- 12 12 0 33991 34177 3.00 6.206 0.13 1 1.60 LC a b 620! 0.00 1839+ 88+ 0.000 62
15:-- 1 1 0 34178 34932 270.00 25.158 0.15 3 2.04 A LC S a 1747= 0.00 1656+ 95+ 0.000 174
26:-- 5 3 2 38680 39279 6.00 19.987 0.52 0 1.60 L S b 4346= 0.00 1803+ 97+ 0.000 434
27:++ 2 2 0 39280 48565 0.07 309.810 0.06 0 1.60 EL s 494= 0.96 1378- 80- 0.000 49
28:++ 14 1 13 48566 52661 0.11 136.637 0.05 0 1.60 EL b 444= 0.90 1032= 70- 0.000 44
29:-- 15 14 1 52662 53572 6.00 30.364 0.07 0 1.60 L S s b 1076+ 0.00 1357- 87+ 0.000 107
I am willing to host a library of logo files. Let me know what kind of information you would like to record with the logo file. I was thinking about the resolution used when recording, the name of the broadcaster, channel number? Amount of border used? Aspect ratio of broadcast ( the logo position often depends on this aspect ratio)
-
drbenjamin
- Posts: 15
- Joined: Sun Dec 09, 2007 5:41 am
logo files
Hi Erik
I hadn't considered that the parameters you mention are relevant; I had focussed on the bitmap stored in the logo files but now I see the coordinates stored at the top of the file. Perhaps a public share of logo files would be unwieldy, given that people likely use different resolutions for different shows (I know I do). I'm a gbPVR user, and I'm not sure that gbPVR makes the resolution of the recording available, so I'd need to hardcode .ini files on a per show rather than a per channel basis.
Your suggestions are valuable, though. What I think I'll do is look through the various logs and find the .logo.txt file that had the highest scores, and see if it scores better when used on diff. recordings.
One other logo question. While doing a bit of tuning on a difficult episode recently, I noticed that the position of the logo had shifted partway through the show, so that the logo was 1/3 outside of the red box. The logo score still seemed pretty good, though; does that mean tha Comskip moves the bitmap mask around to fnd a best fit?
Thanks!
I hadn't considered that the parameters you mention are relevant; I had focussed on the bitmap stored in the logo files but now I see the coordinates stored at the top of the file. Perhaps a public share of logo files would be unwieldy, given that people likely use different resolutions for different shows (I know I do). I'm a gbPVR user, and I'm not sure that gbPVR makes the resolution of the recording available, so I'd need to hardcode .ini files on a per show rather than a per channel basis.
Your suggestions are valuable, though. What I think I'll do is look through the various logs and find the .logo.txt file that had the highest scores, and see if it scores better when used on diff. recordings.
One other logo question. While doing a bit of tuning on a difficult episode recently, I noticed that the position of the logo had shifted partway through the show, so that the logo was 1/3 outside of the red box. The logo score still seemed pretty good, though; does that mean tha Comskip moves the bitmap mask around to fnd a best fit?
Thanks!
The best way to evaluate the performance of a certain logo bitmap is to have
output_framearray=1
when processing the file and the use the .csv file with comskipGUI to fire up the debug window
The yellow line shows the quality of the logo filter. When there is supposed to be logo the line should be at the top and without dropouts. When there is supposed to be no logo the line should be below the threshold you set for logo detection
There is a green line trough the yellow graph. Once you set
logo_filter=4
or any other value then 0 you filter the logo detection over the given amount of frames. Don't use this unless you experience serious detection problems.
What you observe with the shift of the logo and the detection is still working could best be evaluated with the .csv file. You will notice that the quality of the detection has gotten worse but because the detection is a try-state evaluation it is to some extend robust against these shifts. This of course depends to a large extend to the from and content of the logo.
I am considering to set up a Internet repository for logo and ini files anyway. Just to see if I can build such a web application
output_framearray=1
when processing the file and the use the .csv file with comskipGUI to fire up the debug window
The yellow line shows the quality of the logo filter. When there is supposed to be logo the line should be at the top and without dropouts. When there is supposed to be no logo the line should be below the threshold you set for logo detection
There is a green line trough the yellow graph. Once you set
logo_filter=4
or any other value then 0 you filter the logo detection over the given amount of frames. Don't use this unless you experience serious detection problems.
What you observe with the shift of the logo and the detection is still working could best be evaluated with the .csv file. You will notice that the quality of the detection has gotten worse but because the detection is a try-state evaluation it is to some extend robust against these shifts. This of course depends to a large extend to the from and content of the logo.
I am considering to set up a Internet repository for logo and ini files anyway. Just to see if I can build such a web application