chef 3810 Posted August 18, 2021 Posted August 18, 2021 1 hour ago, samuelqwe said: Start times seem to be fine, so it would be interesting to see why it’s going wrong at the end. If possible, could you send me the .bin fingerprint files for season 1? I’d like to perform some tests to find out why it isn’t working. Here are 20 minute .bin files for GOT ep1 and GOT ep2 GOT.ep1.bin GOT.ep2.bin 1
chef 3810 Posted August 18, 2021 Posted August 18, 2021 (edited) I've been randomly auditing episodes in random series, and most of them seem to spot on, to close enough (with in a second). I'm not sure what is wrong with GOT. You'd think with a massive title sequence like that, would be one of the easier series to calculate. Edited August 18, 2021 by chef
samuelqwe 317 Posted August 18, 2021 Posted August 18, 2021 1 minute ago, chef said: I've been randomly auditing episodes in random series, and most of them seem to spot on, to close enough (with in a second). I'm not sure what is wrong with GOT. You'd think with a massive title sequence like that, would be one of the easier series to calculate. There must be something that's interrupting it just long enough to trick the system into thinking that's the end of the intro. I am looking into it.
chef 3810 Posted August 18, 2021 Posted August 18, 2021 Just now, samuelqwe said: There must be something that's interrupting it just long enough to trick the system into thinking that's the end of the intro. I am looking into it. Like using a secondary server on a different machine, and requesting data from drives over a network connection... Type of interruptions?
samuelqwe 317 Posted August 18, 2021 Posted August 18, 2021 1 minute ago, chef said: Like using a secondary server on a different machine, and requesting data from drives over a network connection... Type of interruptions? Sorry I wasn’t clear, I meant in the comparison between the two episodes there must be a slight difference in the intros that is just different for long enough to cause the algorithm to think the intro ended.
rbjtech 5284 Posted August 18, 2021 Posted August 18, 2021 (edited) It does seem odd - the 'HBO' bit is very short (and should be ignored?), but the GoT theme is constant - there are no sound breaks in it .. My test is on a local machine - Season 2 is ok, so it's 'something' about those Season 1 files. I have a different source (4K) from the UHD's - it will take a while to copy them over, but I'm going to see if the same thing happens with those .. they also don't have recaps .. Edited August 18, 2021 by rbjtech
Sammy 790 Posted August 18, 2021 Posted August 18, 2021 For whatever reason I don't have Seasons 1 or 2 and I need to fix that but here's my results for Season 3. Do you need anything other than this graphic?
rbjtech 5284 Posted August 18, 2021 Posted August 18, 2021 (edited) 3 minutes ago, Sammy said: For whatever reason I don't have Seasons 1 or 2 and I need to fix that but here's my results for Season 3. Do you need anything other than this graphic? Do you have a screen grab of Season 1 and what is the source, disk or OTA/Cable (with 'recaps') ? sorry - just re-read your post - ignore me ! Edited August 18, 2021 by rbjtech
Sammy 790 Posted August 18, 2021 Posted August 18, 2021 Source is Sonarr.. IIR most of mine had recaps but it has been a while since I watched this series.
chef 3810 Posted August 18, 2021 Posted August 18, 2021 57 minutes ago, rbjtech said: It does seem odd - the 'HBO' bit is very short (and should be ignored?), but the GoT theme is constant - there are no sound breaks in it .. My test is on a local machine - Season 2 is ok, so it's 'something' about those Season 1 files. I have a different source (4K) from the UHD's - it will take a while to copy them over, but I'm going to see if the same thing happens with those .. they also don't have recaps .. HBO bit will be ignored because its duration is less then the default threshold of 10 seconds. But the plugin will take the first match as being valid. There is the possibility have having more then one match. Right now if episode 1 and episode 2 match we don't bother trying to match them again, but we could try episode 1 and episode 3. Then see if there is a discrepancy. if the results are off, test 1 and 4 to try and resolve any inconsistent results. 2
rbjtech 5284 Posted August 18, 2021 Posted August 18, 2021 (edited) 1 hour ago, rbjtech said: It does seem odd - the 'HBO' bit is very short (and should be ignored?), but the GoT theme is constant - there are no sound breaks in it .. My test is on a local machine - Season 2 is ok, so it's 'something' about those Season 1 files. I have a different source (4K) from the UHD's - it will take a while to copy them over, but I'm going to see if the same thing happens with those .. they also don't have recaps .. So to add - on a fresh 4K UHD source (no recaps) - Season 1 is identified fine - with a IntroStart @ 0:00 and IntroEnd @ 1:43 edit - s01e01 is missing as the cutoff was 10 mins to speed up the fingerprinting .. edit2 - so I added s01e01 back in and it appears to have broken it again. Encoding length set to 30 mins. Why would it break Episode 4 ? It worked perfectly well when Episode 1 was 'skipped'. Something is up a little messed up somewhere. Please let me know if you need logs etc. Thanks Edited August 18, 2021 by rbjtech
Sammy 790 Posted August 18, 2021 Posted August 18, 2021 (edited) I have all of them it seems.. My library is a bit of a mess but Emby seems to get it right.. Whereas the plugin shows this: Edited August 18, 2021 by Sammy 1
samuelqwe 317 Posted August 18, 2021 Posted August 18, 2021 (edited) 19 minutes ago, Sammy said: I have all of them it seems.. My library is a bit of a mess but Emby seems to get it right.. Whereas the plugin shows this: They’re both showing the same thing. If you look closely, Emby is also naming those seasons "Game of Thrones". Also, you seem to have doubles of many seasons, so that might be confusing Emby. @chef I’m still testing with the fingerprints, but I believe my hypothesis is correct. It seems there are bits in the intro that are just different for long enough to cut off the detection early. I don’t know why that’s the case, but it’s what I’m seeing. Edited August 18, 2021 by samuelqwe
chef 3810 Posted August 18, 2021 Posted August 18, 2021 (edited) 1 hour ago, samuelqwe said: @chef I’m still testing with the fingerprints, but I believe my hypothesis is correct. It seems there are bits in the intro that are just different for long enough to cut off the detection early. I don’t know why that’s the case, but it’s what I’m seeing. I wonder how to allow the data to 'not align', if there is a section of data that falls into the contiguous regions later in the arrays? Edited August 18, 2021 by chef
chef 3810 Posted August 18, 2021 Posted August 18, 2021 If you read this computer science pdf, it looks like they add a layer of noise to the audio stream which... for some reason creates a better detection, they also drop the audio quality to 11025. Dunno... IJCSET17-08-05-021.pdf 1
samuelqwe 317 Posted August 18, 2021 Posted August 18, 2021 30 minutes ago, chef said: I wonder how to allow the data to 'not align', if there is a section of data that falls into the contiguous regions later in the arrays? I have an idea on how we could possibly get around this, but we would need to test it to see if it works or causes issues. Essentially, my idea is to look at how long the "interruption" is and if it’s shorter than a certain amount (fixed or based on a percentage of the length of the intro we have detected so far) we simply ignore it and keep going, otherwise stop. 1
samuelqwe 317 Posted August 18, 2021 Posted August 18, 2021 4 minutes ago, chef said: If you read this computer science pdf, it looks like they add a layer of noise to the audio stream which... for some reason creates a better detection, they also drop the audio quality to 11025. Dunno... IJCSET17-08-05-021.pdf 329.63 kB · 0 downloads By lowering the quality, we take away very small difference and only keep the bigger differences, which would probably improve the detection overall. Worth a try for sure. As for adding a layer of noise, I have no idea if it would help us, but we could try it as an experimental setting. I’m sure there’s a way to do that in FFmpeg.
Sammy 790 Posted August 18, 2021 Posted August 18, 2021 3 hours ago, samuelqwe said: They’re both showing the same thing. If you look closely, Emby is also naming those seasons "Game of Thrones". Also, you seem to have doubles of many seasons, so that might be confusing Emby. Well! I only looked at the top, not the bottom..
samuelqwe 317 Posted August 18, 2021 Posted August 18, 2021 3 hours ago, samuelqwe said: I have an idea on how we could possibly get around this, but we would need to test it to see if it works or causes issues. Essentially, my idea is to look at how long the "interruption" is and if it’s shorter than a certain amount (fixed or based on a percentage of the length of the intro we have detected so far) we simply ignore it and keep going, otherwise stop. As it turns out, we already do this, but maybe we can play around with how long it should wait until it actually stops. After further testing with episodes from season 1 of GOT, I've found that it works fine for the majority of episodes, except the first two. As long as you don't compare with the first two, everything seems to line up. I don't know why those two specific episodes cause issues, but they do. 5 hours ago, chef said: Right now if episode 1 and episode 2 match we don't bother trying to match them again, but we could try episode 1 and episode 3. Then see if there is a discrepancy. if the results are off, test 1 and 4 to try and resolve any inconsistent results. This would likely help with GOT in this case. Perhaps lowering the sample rate further, as you suggested, would also help in this situation.
chef 3810 Posted August 18, 2021 Posted August 18, 2021 (edited) 18 minutes ago, samuelqwe said: As it turns out, we already do this, but maybe we can play around with how long it should wait until it actually stops. After further testing with episodes from season 1 of GOT, I've found that it works fine for the majority of episodes, except the first two. As long as you don't compare with the first two, everything seems to line up. I don't know why those two specific episodes cause issues, but they do. This would likely help with GOT in this case. Perhaps lowering the sample rate further, as you suggested, would also help in this situation. I'll try it out now. @samuelqwe, here is a crude version of the plugin. It'll only run on Linux currently, or until Emby adds the ffmpeg controls for chromaprint. Sammy and rbjtech if you want to try it on windows, just be sure to back up Emby's ffmpeg, and I'll PM you a substitute version for windows with the proper commands. NOT FOR WINDOWS : IntroSkip_v2.zip same old story, clear browser history . Edited August 19, 2021 by chef 3
rbjtech 5284 Posted August 19, 2021 Posted August 19, 2021 Thanks guys - will try this out today. Your results are similar to mine when using the 4K 'clean' source (remux). Post All is well with Season 1 with the exception of Episode 1, which then throws a spanner in the works if you re-add it. I'm going to experiment today (copying Ep1 as Ep2, 3, 4 etc to see if it then picks it up properly.) Also - which Audio track is it using (presumably the default one) and is it doing a channel 'mix' and then analysis a mono track ? In other words, does the Audio track type make any difference to what is being analysed ?
rbjtech 5284 Posted August 19, 2021 Posted August 19, 2021 (edited) So I think I see what is going on - I've created my own S1 episodes from the 4K Remux (very low Video quality, but Stereo AAC from the THD Atmos Track) - I did to ensure I have a solid baseline and also to speed up the process. I believe the HBO part IS important - as Ep1 - You get the HBO at the start (8 seconds), then you get the Intro ~7 minutes in but with no extra HBO. All the other episodes have HBO as part of the Intro, so Ep1 is NOT the same as all the other episodes.. I'm going to edit Ep1 to put the HBO back in just before the real Intro - and I'm betting it then picks it back up and with the correct timings like Ep2,3,4 etc Edited August 19, 2021 by rbjtech
chef 3810 Posted August 19, 2021 Posted August 19, 2021 I'm testing some new code. If the seasons episode count is big enough, try comparing the episode (with no sequence data) with more then one other episode, and check the results against each other to see if there is a large discrepancy in the result. It will cause longer scan times, but is worth getting the best data possible.
rbjtech 5284 Posted August 19, 2021 Posted August 19, 2021 I don't think the code is the problem - it is doing exactly as it was designed to do. In summary Ep1 is not being picked up as having an Intro - which is 'correct' as it's intro does not match the other episodes. I'm just proving the point by inserting the full Intro into Ep1 - but having trouble with matching the I-frames as it didn't like my previous attempt - so I assume it has to be bit-perfect (ie not an 'insert' ) 1
chef 3810 Posted August 19, 2021 Posted August 19, 2021 (edited) Should the table be editable then? I know there was concern before about having concrete data that a user wouldn't have to alter, but if 90% of the data is good data, maybe allowing the user the ability to alter the rest is okay? Of course, it does defeat the purpose of watching a show for the first time and having the option to skip an intro be only 90% of the time properly displayed. How does Plex do it? Is their data editable? Is their data perfect? Edited August 19, 2021 by chef
Recommended Posts