MCTemporalDenoise [v1.4.20 - Update 2010/07/02]

[tormento]

Originally Posted by Didée

There already is one. (Without SC detection, of course.)

I was talking about MCTD.

Don't know why but my "porting" to x64, disabling Sangnom, is awfully slow.

[asarian]

Originally Posted by asarian

LaTo, I may not be doing it entirely right yet. :) I'm using this avs script on a 1-minute clip:

LoadPlugin("C:\Program Files\AviSynth 2.5\plugins\FFMS2.dll")
Loadplugin("C:\Program Files\AviSynth 2.5\plugins\FFT3DFilter.dll")
Loadplugin("C:\Program Files\AviSynth 2.5\plugins\mt_masktools-25.dll")
Loadplugin("C:\Program Files\AviSynth 2.5\plugins\mvtools2.dll")
Loadplugin("C:\Program Files\AviSynth 2.5\plugins\RemoveGrainSSE3.dll")
Loadplugin("C:\Program Files\AviSynth 2.5\plugins\TTempSmooth.dll")
Loadplugin("C:\Program Files\AviSynth 2.5\plugins\RepairSSE3.dll")
Loadplugin("C:\Program Files\AviSynth 2.5\plugins\deblock.dll")
Loadplugin("C:\Program Files\AviSynth 2.5\plugins\DctFilter.dll")
Loadplugin("C:\Program Files\AviSynth 2.5\plugins\EEDI2.dll")
Loadplugin("C:\Program Files\AviSynth 2.5\plugins\SangNom.dll")
Loadplugin("C:\Program Files\AviSynth 2.5\plugins\gradfun2db.dll")
Loadplugin("C:\Program Files\AviSynth 2.5\plugins\AddGrainC.dll")

video=FFVideoSource("M:\jobs\2010test.mkv").ConvertToYV12()
video=MCTemporalDenoise(video, settings="very high")
video=Crop(video, 0, 142, 0, -142)

return video

Nothing happens for a long while, though. First I get a Windoze error, saying it can't use "mt_masktools-25-x64.dll" (of course it can't: I'm on 32bit; and I didn't ask for it to be run, either). Then x264 keeps hovering around 25% CPU, steadily increasing in memory size (to around 1.4GB). And finally it tells me that 1 minute clip, at 0.00fps, (sic!) will take over 300 hours! Why, that seems a mite extreme. :)

x264 remains at ca. 25% CPU. So, I guess I'm doing something wrong. Any hints?

Anyone? Anyone? Bueller? Bueller? :)

Seriously; I may not have the fastest machine (Q6700 @ 3.6 Ghz); but 300 hours for a 1 minute clip??! That's looking at a ~4 year encode!!

I could probably crop a little more efficiently; like first do 400, -400, so at to still be div 16 (one of the filters was complaining about that), and only then do MCTemporalDenoise, and only then crop an extra 2, -2 afterwards:

video=FFVideoSource("M:\jobs\2010test.mkv").ConvertToYV12()
video=Crop(video, 0, 140, 0, -140)
video=MCTemporalDenoise(video, settings="very high")
video=Crop(video, 0, 2, 0, -2)

But still, these encoding times seem too long (and unworkable; I mean, who can keep his machine on for 4 years? Hoping it won't reset in the meantime).

[Gavino]

Originally Posted by asarian

I get a Windoze error, saying it can't use "mt_masktools-25-x64.dll" (of course it can't: I'm on 32bit; and I didn't ask for it to be run, either).

Is mt_masktools-25-x64.dll in your plugins folder (C:\Program Files\AviSynth 2.5\plugins)? If so, it will be loaded automatically by Avisynth, so you want to remove it from there.

For the same reason, all your LoadPlugin commands are unnecessary.

[asarian]

Originally Posted by Gavino

Is mt_masktools-25-x64.dll in your plugins folder (C:\Program Files\AviSynth 2.5\plugins)? If so, it will be loaded automatically by Avisynth, so you want to remove it from there.

For the same reason, all your LoadPlugin commands are unnecessary.

Yeah, I put mt_masktools-25-x64.dll there recently, as part of upgrading all required tools for MCTemporalDenoise. Didn't realize they all autoload. :)

P.S. Any idea why x264 constantly only takes about 25% CPU for MCTemporalDenoise? That doesn't exactly help encoding times either. :(

[LaTo]

Originally Posted by asarian

Yeah, I put mt_masktools-25-x64.dll there recently, as part of upgrading all required tools for MCTemporalDenoise. Didn't realize they all autoload.

P.S. Any idea why x264 constantly only takes about 25% CPU for MCTemporalDenoise? That doesn't exactly help encoding times either.

x264 wait a long time after MCTD... Because avisynth isn't multithreaded, MCTD use only one core of your quad (25%).

[asarian]

Originally Posted by LaTo

x264 wait a long time after MCTD... Because avisynth isn't multithreaded, MCTD use only one core of your quad (25%).

LaTo, you're my new hero!! :)

I just tried my sample with 'settings=high', and instead of taking 300 hours (for 'very high'), it was done in 30 minutes! :) Which is very reasonable. And the result? Not only does MCTemporalDenoise pass the SC test (whereas TemporalDegrain, on the defaults, failed), but the resulant video is spendid! Beyond belief, so good actually!!

I can't get 'GPU=true' to work yet (GTX 260), though:

avs [info]: 1920x796p 1:1 @ 24000/1001 fps (cfr)
x264 [info]: using SAR=1/1
x264 [info]: using cpu capabilities: MMX2 SSE2Fast SSSE3 Cache64
x264 [info]: profile High, level 4.1
x264 [error]: malloc of size 6856704 failed/s, eta 19:40:54
x264 [error]: x264_encoder_encode failed

Not sure what that is all about, but, regardless, I'm already happy. :)

[LaTo]

v1.4.20 :
tweaked some settings
updated QED_Deblock version

[Boulder]

Thanks

Are you by any chance planning on doing a quick-and-dirty fix for the missing scenechange detection?

[LaTo]

Originally Posted by Boulder

Thanks

Are you by any chance planning on doing a quick-and-dirty fix for the missing scenechange detection?

Yes it's on my todo list!

[asarian]

When I run MCTemporalDenoise with 'enhance=true' ("medium"), suddenly a GHOST PORTRAIT appears in one of the frames in my Blade Runner encode. Oops! :) When I add a GradFun2DBmod command after MCTemporalDenoise, the issue is resolved again.

Maybe this needs a little looking into?

[LaTo]

Originally Posted by asarian

When I run MCTemporalDenoise with 'enhance=true' ("medium"), suddenly a GHOST PORTRAIT appears in one of the frames in my Blade Runner encode. Oops! When I add a GradFun2DBmod command after MCTemporalDenoise, the issue is resolved again.

Maybe this needs a little looking into?

It's not a problem with enhance=true because it's only spatial.
Do you use DirectShowSource() or GPU=true in your script?

[asarian]

Originally Posted by LaTo

It's not a problem with enhance=true because it's only spatial.
Do you use DirectShowSource() or GPU=true in your script?

No, I always use FFVideoSource for frame consistency.

I did use 'GPU=true', though. How would GPU usage cause just a temporal displacement, though?

EDIT: I *do* stream using avs2yuv now, though, because otherwise x264 would run out of memory. Could using avs2yuv to stream make things frame inconsistent again?

[LaTo]

Originally Posted by asarian

No, I always use FFVideoSource for frame consistency.

I did use 'GPU=true', though. How would GPU usage cause just a temporal displacement, though?

EDIT: I *do* stream using avs2yuv now, though, because otherwise x264 would run out of memory. Coould using avs2yuv to stream make things frame inconsistent again?

GPU=true causes ghosting when seeking, moreover I have somes issues with HD-1080p.
So GPU=false is the default

For avs2yuv, I don't know... Maybe, I never use it.

[asarian]

Originally Posted by LaTo

GPU=true causes ghosting when seeking, moreover I have somes issues with HD-1080p.
So GPU=false is the default :D

For avs2yuv, I don't know... Maybe, I never use it.

Well, that's it for "GPU=true" then. :) Wasn't getting a speed improvement out of it anyway; just a memory benefit.

The avs2yuv frame consistency matter has got me worried a bit (if it is, indeed, an issue). Maybe someone else knows?

[asarian]

Okay, some very weird stuff going on. I used this following line for x264 on this Total Recall TEST SCENE:

"C:\x264\avs2yuv.exe" "M:\jobs\trecalltest.avs" -o - | "C:\x264\x264.exe" - --demuxer y4m --crf 14 --sar 1:1 --aud --nal-hrd none --level 4.1 --preset placebo --vbv-bufsize 70000 --vbv-maxrate 60000 --bframes 8 --merange 32 --aq-mode 2 --ref 4 --tune grain --output "C:\video\trecalltest.264"

And the AVS scripts is as follows:

FFVideoSource("M:\jobs\trecalltest.mkv").ConvertToYV12()
Crop(0, 4, 0, -4)
MCTemporalDenoise(settings="medium", enhance=true)
Crop(0, 20, 0, -20)

The ghosting/misplaced frames in the RESULT are numerous. :( What the heck is going on!? The weird part is, that I can get this to work properly when I vary CRF and qpmin some. For instance, CRF = 10 + QPMIN = 10 gives me a flawless encode. This faulting file, however (CRF = 14 + QPMIN = 10) totally flips out. You'd think that makes it a x264 problem; but not so! If I run just the AVS script alone, and watch the output in GOM Player (very slowly, of course), I get some frame mess-ups too.

So, it seems to be a timing problem of some sort rather (in which x264 settings, only indirectly, play a role). Could someone test the above clean sample with the same x24 settings I used, please? Very mysterious things going on. :)

[asarian]

Originally Posted by LaTo

It's not a problem with enhance=true because it's only spatial.

LaTo, I'm really thinking it may be 'enhance=true' after all. When I remove the enhance option from the MCTemporalDenoise line, and export it simply as a second post-filter, everything goes well again!

FFVideoSource("M:\jobs\trecalltest.mkv").ConvertToYV12()
Crop(0, 4, 0, -4)
MCTemporalDenoise(settings="medium")
GradFun2DBmod(thr=1.4,thrC=1.0,mode=0,str=1.2,strC=0.0,temp=50,adapt=64,mask=false,show=false)
Crop(0, 20, 0, -20)

This script, contrary to the other, yields a perfect encode, under any conditions! So, maybe the 'in-house' enhance feature from MCTemporalDenoise corrupts something internally on occasion (like a buffer accidentally being overwritten? Or some variables maybe?). Whatever the case may be, doing GradFun2DBmod seperately definitely solves the issue, on every test I've run so far.

[LaTo]

And with temp=-1 ..?

Which version of avisynth do you use?

[asarian]

Originally Posted by LaTo

And with temp=-1 ..?

No, that makes no difference: same messed-up output.

Which version of avisynth do you use?

I'm using regular AviSynth version 2.5.8 (non-MT!). Why? You think it's a threads issue somehow?

Anyway, thanks for looking into it.

[LaTo]

Really strange.... I will try with your sample+script.

[asarian]

Originally Posted by LaTo

Really strange.... I will try with your sample+script.

Thanks, appreciate it. :)

In the meantime at least I can start a 'sane' encode again, using GradFun2DBmod on a separate line.