New Free Msu Deinterlacer

[diablo23q]

Hello, everyone!
I'm proud to present new MSU VirtualDub filter
Here it is - http://compression.ru/video/deinterlacer/index_en.html
Report bugs to deinterlacer _at_ graphics.cs.msu.ru

[Selur]

just wondering if they used the Avisynth MT or the normal Avisynth Version for their tests since that could really change a lot for the speed/quality comparison table,...

[cord-factor]

Looks like your filter is pretty good. Any hopes that it will be ported to ffmpeg?

[kolak]

Hmmm- I tried it on these clips:

http://www.avsforum.com/avs-vb/showthread.php?t=1157287

and it's not as good as QTGMC.

It also crashing on HD files.

[Taurus]

and it's not as good as QTGMC.
Maybe I'm blind, but could not find any settings they used for the compare...
Maybe QTGMC(superduperfast)?
But anyway I will give it a try on some true interlaced source.

[Gser]

lol PSNR graphs. Tells about as much about picture quality as asking a blind person. As you can see McBOB scored higher than QTMC in all the grahps yet it produces inferior quality. Not to mention they failed to mention what settings they used with QTMC. Perhaps this will be a good replacement for YADIF.

[kolak]

It may be as good as YADIF, but it's no near QTGMC.
I think it uses many frames as reference, but it's bit buggy.

[kolak]

and it's not as good as QTGMC.
Maybe I'm blind, but could not find any settings they used for the compare...
Maybe QTGMC(superduperfast)?
But anyway I will give it a try on some true interlaced source.

This is quite good test:

http://dl.dropbox.com/u/8424544/Slices/Video/SD_NTSC_29.97_480i.ts

[CruNcher]

Get the full report they also measured with SSIM, and their it even beats nnedi3 most of the time :D :)
diablo23q is this the deinterlacer you worked on with Samsung Research ?

I really wonder how it compares to Intels ASIC Deinterlacer :)

[kolak]

Get the full report they also measured with SSIM, and their it even beats nnedi3 most of the time :D :)
diablo23q is this the deinterlacer you worked on with Samsung Research ?

I really wonder how it compares to Intels ASIC Deinterlacer :)

Try it on real footage- it's not that great at all.

[CruNcher]

yes they also explain that it depends on the interpolation of the source if it isn't raw it depends which resizer was used for interpolation if the deinterlacer and resize algorithm match the result will be better, this could also really well explain some of my Intel ASIC results if they also use the same interpolation in their Deinterlacer as in their Resizer it might explain the better quality vs Yadiff in my tests, which was Realtime upscaling from SD to HD with Intels Propiertary Lanczos4 Scaler.


During the comparison it was noticed that some transformations of
progressive video may affect comparison results. For example, similarity of
interpolation algorithms in resizing and deinterlacing leads to higher quality
metric value for this deinterlacer. Relation between scaling factor and PSNR
value is shown on these charts for Lanczos and bicubic resizing. PSNR value
in the point where scaling factor equals 1 stands out of common chart
behavior. This fact proves an importance of correct source videos in this kind
of comparisons. Processing a video that has been converted from resized
progressive source is a very rare usecase for deinterlacing algorithms, so our
comparison was performed only with unscaled test sequences.

[Didée]

Realworld source: DVB-S capture, PAL 576i

MSU Deint v2.0 versus TGMC: <Klick!> (http://www.mediafire.com/?zfd126r6a2x1fe0) (~30MB @ MediaFire)


Now, what to say? I'm speechless!

[kypec]

Now, what to say? I'm speechless!
Yeah, me too! :goodpost: MSU deinterlaced picture is so... live... that it can be hardly watched. :D

[CruNcher]

Indeed that looks yadiff comparable on the first sight though according to metrics it should be much better ;)
Though rating it only with 1 Scene result isn't fair, though this shimmering is a big perceptible problem indeed and a good Deinterlacer shouldn't fail with it ;)
Except on duck take off their needi3 shows better metric performance i wonder if we would see shimmering here as a result of this lower metric result :)

i think they are 4 key areas to rate independently

1. Shimmering Prevention (MSU Fail)
2. Detail Reconstruction (Sharpness)
3. Motion Stability
4. Speed

and i would say for live results in that order or if you prefer switching

1. Shimmering Prevention (MSU Fail)
2. Motion Stability
3. Detail Reconstruction (Sharpness)
4. Speed

and for realtime results

1. Speed
2. Shimmering Prevention (MSU Fail)
3. Motion Stability
4. Detail Reconstruction (Sharpness)

or the other way too

1. Speed
2. Shimmering Prevention (MSU Fail)
3. Detail Reconstruction (Sharpness)
4. Motion Stability


For Intel (Realtime) Shimmering prevention is really good where MSU and Yadiff fail (though it includes upscaling to HD with Intels Scaler)

This is quite good test:

http://dl.dropbox.com/u/8424544/Slices/Video/SD_NTSC_29.97_480i.ts

Most recognizable visual issue in this Intel SD->HD Realtime chain is left aliasing @ the Edge of the Glases

Yadiff flickers in the test card and also shows more perceptible problem @ the Edges

[Terka]

how you were able to run MSU deinterlacer? im unable to use it under Vdub 1.9.11.

[Taurus]

how you were able to run MSU deinterlacer? im unable to use it under Vdub 1.9.11.
Maybe you do not have a SSE2 compatible processor :rolleyes:;)

[Terka]

Intel i5 mobile?

[kolak]

I have way worse processor in my laptop and it works fine.

Get latest Vdub.

It crashes on every machine for me on HD source.

[Terka]

1.9.11 is the latest stable version.
when i put the dll to plugins directory, vdub will not load it.
when loaded manually, using "load" button, error appears:
Cannot load plugin module "C:\VirtualDub-1.9.2\plugins\MSU_Deinterlacer.vdf": This application has failed to start because the application configuration is incorrect. Reinstalling the application may fix this problem.

[LoRd_MuldeR]

Sounds like a missing DLL dependency. Do you have the Visual C++ 2008 Redistribute Package (32-Bit) (http://www.microsoft.com/download/en/details.aspx?id=29) installed?

Also: You are not trying to use a 32-Bit plug-in DLL with 64-Bit VirtualDub or vice versa, right? ;)

[LoRd_MuldeR]

Seems like MSU Deinterlacer has some problem with chroma processing.

There are some annoying chroma interlacing artifacts in the output:
http://img638.imageshack.us/img638/8224/msuvsyadif.th.png (http://img638.imageshack.us/img638/8224/msuvsyadif.png)

Other than that, it seems to produce more "stable" output than plain Yadif.

[Mounir]

I'm not impressed thus far, it creates weird artifacts where there should be none imo

[kolak]

Seems like MSU Deinterlacer has some problem with chroma processing.

There are some annoying chroma interlacing artifacts in the output:
http://img638.imageshack.us/img638/8224/msuvsyadif.th.png (http://img638.imageshack.us/img638/8224/msuvsyadif.png)

Other than that, it seems to produce more "stable" output than plain Yadif.

Yep- I think it needs some update/bug fixes and than it's not bad.

[diablo23q]

Hello everyone! Thanks for your responses. I totally agree about PSNR graph, it's not representative.
For QTGMC 'slow' preset was used. And MT Avisynth version.
As for shimmering, honestly, i didn't expect that, looks very bad :(
Chroma interlacing is strange. LoRd_MuldeR, could you please tell colorspace of source file? It'll be great if you'll send me this file.
HD crash - fixed.
So, this filter will be re-released and bugfixed, wait for new versions, thank you for your patience.

[LoRd_MuldeR]

This was actually my sample. I can upload the file somewhere and send you the link via PM.

(But it may take some time, as my internet connection is not very fast)

[kolak]

Hello everyone! Thanks for your responses. I totally agree about PSNR graph, it's not representative.
For QTGMC 'slow' preset was used. And MT Avisynth version.
As for shimmering, honestly, i didn't expect that, looks very bad :(
Chroma interlacing is strange. kolak, could you please tell colorspace of source file? It'll be great if you'll send me this file.
HD crash - fixed.
So, this filter will be re-released and bugfixed, wait for new versions, thank you for your patience.

It's not me who showed chroma issues. I had HD crash- I use mainly YUY2.

[kolak]

Any news- eta for new version?