TempGaussMC showing jaggy diagonal lines

[Dave2ic]

My hard drive failed a few days ago and I had to reinstall Windows and download and setup all my AviSynth scripts again. However, this time when I ran TempGaussMC_beta1, jaggy edges in diagonal lines showed in many frames that I’m sure weren’t there when I previously ran it in my previous installation. I assume this must be something to do with the versions of some of the plugins or differences in my script – although the script is about as basic as it can be:

AVISource("e:\Test.avi")
ConvertToYV12()
TempGaussMC_beta1(edimode="nnedi2").SelectEven()

Versions of the required plugins are as follows:

MVTools (mvtools.dll version is 1.11.4.5)
RemoveGrain 0.9
MaskTools v2 (mt_masktools-25.dll version is 2.036)
NNEDI2 (nnedi2.dll version is 1.4)

Does this all look ok? It’s been difficult to remember exactly where I downloaded all the different plugins so I can’t be sure I have AVISynth setup exactly as I did before.

To illustrate what I mean, please see the following image. It shows 100% crops of the same frame of the various deinterlacing filters I have tried:





You can see that the white line I have circled on the bottom right image is much jaggier using the TempGaussMC filter than the NNEDI filter. TempGaussMC also runs nearly twice as fast as before, so there must have been something different about my previous setup.

Also, I find I get some odd pixels showing up in lots of frames in the NNEDI2+YADIFmod filter – see the circle in the top-right image as an example. Is this normal? Going by my tests I get better results with NNEDI2 on its own than NNEDI2+YADIFmod or TempGaussMC - although I'd prefer TempGaussMC if I can get these jaggy lines sorted out.

Any advice/comments greatly appreciated.

[poisondeathray]

You forgot converttoyv12(interlaced=true)

If that's not enough, try changing the edimode of TGMC to default (eedi2)

I find I get some odd pixels showing up in lots of frames in the NNEDI2+YADIFmod filter – see the circle in the top-right image as an example. Is this normal?

It can happen, I've noticed this too. But NNEDI2 by itself can drop lines along horizontal objects like signs and letters , so it's not perfect either.

You will find every deinterlacer will have a weakness on certain scenes or types of content. You might want to process certain segments with different filters

I've noticed EEDI3 is great on those diagonal lines such as tennis court lines, or football field lines, but it creates horrible artifacts on other objects . You can see great examples of this in the NNEDI2 development thread at Doom9. The only way to get it nearly "perfect" is to mask out certain parts of the frame using different filters - but this is very tedious to do.

Can you post the a sample of the section you are trying to process?

[jagabo]

Yes, you're discovering that there is no perfect deinterlacer. In fact, there is no perfect way to deinterlace, even theoretically.

If you post a short clip others can try out the deinterlacers and let you know if they get same thing.

[Dave2ic]

If that's not enough, try changing the edimode of TGMC to default (eedi2)

That did the trick! Thanks for your prompt help (again!) poinsondeathray.

Here is the result:



So using EEDI2 resolved the jaggy line problem, but it is insanely slow! I only get 0.16 fps using TempGauss with EEDI2, whereas I get 0.93 fps using TempGauss with NNEDI2, and 2.8 fps with NNEDI2 alone!

For my purposes the best option seems to be using NNEDI2 alone to deinterlace. I haven't noticed the potiential problems with 'dropped lines' you refer to with NNEDI2 in my videos - probably because it's handheld camcorder material (so the camera is always moving) and there are no logos or subtitles to deinterlace.

Thanks again for your help.

[Dave2ic]

Yes, you're discovering that there is no perfect deinterlacer. In fact, there is no perfect way to deinterlace, even theoretically.

Actually, they all do a great job when you compare them to the horrible results I got with Smart Deinterlace in VirtualDub. And the clips I am testing are areas of pretty high speed motion, so you don't really even notice the difference played at full speed. Having said that, if I'm going to be going through the trouble of deinterlacing it's still worthwhile evaluating the various options.

For my handheld camcorder video they all look fine played at full speed. But I could imagine deinterlacing it becomes more challenging if you have, say, a tennis match with a combination of court lines (as referred to by poisondeathray), fast motion, subtitles/logos and a steady camera (as opposed to a handheld that is moving all the time).

[Dave2ic]

Just a further note to my deinterlacing tests (in case someone finds them of value!).

While I was happy with NNEDI2 on its own with the testing I had done, after comments by poisondeathray that it can have problems of its own by dropping horizontal lines I decided I should test for detail retention when the camera is held still – not just with the handheld camcorder video I had been using up until now.

This showed some interesting results as shown in the image below.




This is a 100% crop of my father tying his shoe laces, so there’s deinterlacing involved with the moving hand, but it also gives me a chance to test for detail retention in the rest of the image, particularly the cane chair. And this is where NNEDI2 on its own really does show its shortcomings. As you can see where I’ve circled in red, it really does make a mess of the cane chair detail compared to the original. TempGaussMC and TempGaussMCMod have no such problems.

The reason I have tried TempGaussMCMod here refers back to the orginal post – the jaggy diagonal lines I was getting. I tried the implementation of TempGaussMC that I found at the following site, and experienced no such problems:

http://in-trans.appspot.com/entry/fictionjunction-yuuka-popjam-2005

Just scroll down to the bottom and there’s the script and handy links to all the plugins used (I left out the ‘mpeg2source line’ of course and the tr2=1 (wasn’t sure what that was for)).

I’m very pleased with the results of this TempGaussMCMod script with the modules list on this page, so this will be my deinterlacer of choice now, given that I’m unlikely to need to process more that 5-10 minutes of my own edited video at a time. It’s obviously not practical to use TempGaussMC if you’ve got hours of video to deinterlace.

[poisondeathray]

the beta1_mod is modified to have parts of it use multithreading (i think mvtools branch; - it's still very slow because other parts of the processing are still single threaded)

tr2 controls the radius of the mvdegrain post processing

it's not readily apparent in your particular screenshots , but TGMC denoises and blurs a bit because it was developed specfically handle interline twitter (shimmering). It's a side effect. If you compare other parts like faces etc, or if you have higher quality footage to test on that exibits fine detail... you will notice those processed by TGMC have quite a bit less detail

[Dave2ic]

it's not readily apparent in your particular screenshots , but TGMC denoises and blurs a bit because it was developed specfically handle interline twitter (shimmering). It's a side effect. If you compare other parts like faces etc, or if you have higher quality footage to test on that exibits fine detail... you will notice those processed by TGMC have quite a bit less detail

I know what you are saying - I've read in other places too that TempGaussMC does lose some fine detail. But I haven't experienced that in the video I have tested from my camcorder. I chose the clip above to test for that very thing - to see if the fine detail in the cane chair was retained after deinterlacing - and I was very impressed with how TempGaussMC performed compared to the mess NNEDI2 made of the detail.

[citanool]

Curious to see how this might compare to your preference of TempGauss w/EEDI results: yadifmod(mode=1, edeint=nnedi(field=-2))