ghosting with MMX optimized KernelDeint 1.5.1

[scrtagt69]

Im using the below script here for all my ntsc dvds which follow the 3 progessive 2 interlaced frame patterns, and most of the ones i do you can ghosts. has anoyone else had this problem with the kernel bob. The ghosts arent very noticeable but they are there. and my dvd sources are clean, free of ghosts. any input in this would be helpful.
thank you.


MPEG2Source("test.d2v",ipp=true,cpu=6,idct=7)

DeDot(30,30,10,5)
DeGrainMedian(limitY=5,limitUV=10,mode=0)

KernelBob(order=1,linked=false,sharp=true,threshold=10)
BlendBob(derainbow=1)
Decimate(mode=1,threshold=1.0)

[Mug Funky]

blendbob will blur things slightly, but without an example of what you mean i can't say more than that..

i will try your script on the only 3:2 material i have (PAL land)

[edit]

decimate is the culprit here, not kernelbob.

i used the default values fer decimate and the ghosting went away.

btw, in my opinion, blendbob is a last resort for IVTC - if you have clean sources then either smartdecimate or telecide should suffice (i actually use a combination of both at once for difficult clips, but occasionally still get an orphaned field at scenechanges).

[edit edit]

lol! even blendbob can't handle the orphaned field at scenechage trick. it fails better than other IVTCs i've tried though. i could kick Gainax in the face for doing such a terrible job with the "renewal" Eva editions.

[scharfis_brain]

but occasionally still get an orphaned field at scenechanges

that's what the bob-feature of smartdecimate is for.

[scrtagt69]

so a good recommendation for doing IVTC is smartdecimate? i actually had trouble using that(kept getting errors), i will give it another shot and see what happens. im still interested in another peoples results though with kernel bob.

[Chainmax]

Originally posted by Mug Funky
btw, in my opinion, blendbob is a last resort for IVTC
So, you never tried TMPGEnc's automatic or manual IVTC?

Originally posted by Mug Funky
if you have clean sources then either smartdecimate or telecide should suffice (i actually use a combination of both at once for difficult clips
What kind of combination?

[JuanC]

Originally posted by scrtagt69
MPEG2Source("test.d2v",ipp=true,cpu=6,idct=7)

DeDot(30,30,10,5)
DeGrainMedian(limitY=5,limitUV=10,mode=0)

KernelBob(order=1,linked=false,sharp=true,threshold=10)
BlendBob(derainbow=1)
Decimate(mode=1,threshold=1.0) If you are smoothing / denoiseing before deinterlacing/ ivtc'ing / bobing, shouldn't you tell DeGrainMedian to treat your clip as interlaced? (interlaced=true). Also, does DeDot treat interlaced material appropiatelly?

[kassandro]

Originally posted by scrtagt69
Im using the below script here for all my ntsc dvds which follow the 3 progessive 2 interlaced frame patterns, and most of the ones i do you can ghosts. has anoyone else had this problem with the kernel bob. The ghosts arent very noticeable but they are there. and my dvd sources are clean, free of ghosts. any input in this would be helpful.
thank you.


MPEG2Source("test.d2v",ipp=true,cpu=6,idct=7)

DeDot(30,30,10,5)
DeGrainMedian(limitY=5,limitUV=10,mode=0)

KernelBob(order=1,linked=false,sharp=true,threshold=10)
BlendBob(derainbow=1)
Decimate(mode=1,threshold=1.0)

I think that you are falsely blaming Blendbob for the ghosts. The ghosts are created by the very wrong usage of DeGrainMedian. After all DeGrainMedian is primarily a spatial denoiser and you should NEVER apply a spatial denoiser (such filters are for progressive material only) before deinterlacing or ivtc. It is not as bad as resizing before deinterlacing, because a spatial denoiser deinterlaces interlaced material to a certain extend and in doing so it creates ghosts. If you nevertheless want to apply such filters before deinterlacing, you have to separate the fields.

edit:Sorry JuanC, for mostly repeating your statement, but I didn't refresh the page.

[Mug Funky]

Originally posted by Chainmax
What kind of combination?

i feed smartdecimate's "bob" with a telecide fallback clip like so:


function DoubleTelecide(clip c, bool "post")
{

post = default(post,true)

order = (c.getparity==true)? 1 : 0

odd = (post==true)? c.doubleweave().selectodd().telecide(order=order,guide=1,hints=true).kerneldeint(order=0,threshold=3) :
\ c.doubleweave().selectodd().telecide(order=1-order,guide=1,post=0)

even = (post==true)? c.telecide(order=order,guide=1,hints=true).kerneldeint(order=1,threshold=3) :
\ c.telecide(order=order,guide=1,post=0)

return interleave(even,odd)

}

mpeg2source("blah.d2v")

assumetff()
#assumebff()

b=last.doubletelecide()
smartdecimate(24,60,bob=b)


this works most of the time, but the orphan fields still get deinterlaced (dammit).

@ scharfi: i know what bob is for :) i like to experiment.

though it seems smartdecimate doesn't pay as much attention to chroma as it could. i haven't got to tweaking Tmax yet, because after reading the docs i got kinda scared, and need to get some gumption up before i delve into logfiles and such.

[edit]

it'd be good to have blendbob as an automatic fallback for the above, just for the 1 or 2 times it's needed... but i've no idea how to detect these very few instances. seeing as they're nearly always just after a scenechange, i don't think it really matters.

the above script was basically my hack to stop smartdecimate from doing too much bobbing (it bobs the hell out of still-but-slightly-shaky scenes, no matter what i set "tel" and "noise" to).

[scharfis_brain]

though it seems smartdecimate doesn't pay as much attention to chroma as it could.

smartdecimate is not able to handle YV12 properly.
but using converttoyuy2(interlaced=true) before isn't that good, cause it hurts chroma...

so, whatI really really need are two colorspace conversion routines, that allow lossless YV12 -> YUY2-> YV12 conversion!

this can be achieved by simply using nearest neighbor chroma-picking, instead of the cpu-consuming pixel-weighting for up & down scaling.

this means: converttoyuy2_blah() will just duplicate missing chroma-lines using either progressive or interlaced chroma-line-doubling, depending on how theinterlaced-flag is set to.

also theother wayaround: converttoyv12_blah() it just should remove every 2nd chroma-line out of the stream, maybe it should be customizable, whether the odd or even lines get discharged...

[Mug Funky]

scharfi: can smartdecimate be fed an analysis clip in a different colourspace to the clip it works on? i know it can work with clips of different sizes (like one with cropped-out subs to not interfere with the matching process).

i'd try it, but i'm doing an encoding :)

[scharfis_brain]

it doesn't matter how you feed it.
it just cannot properly handle YV12, which is making it bad, if your source is hard-telecined YV12...

[ObiKenobi]

Originally posted by Mug Funky i could kick Gainax in the face for doing such a terrible job with the "renewal" Eva editions. [/B]

I'm not sure what you mean but I was able to IVTC ADV's platinum dvds perfectly with no combing at all using Decomb521. Scene changes came out perfect.

[Mug Funky]

@ obiwan: look harder :)

it's much better than the original, but the editing was done post-telecine, and seems to have been done on the field level, so there's orphan fields. most notably the intro sequence - the title overlays are 30i interlaced in places (can you believe that?), and frame 2458 (i think it'll vary between episodes) at 29.97 gives headaches to any IVTC - there's a black field between the brown-beige sketch and the japanese characters that i can't read :)

also, the hard telecine means it's absolutely crawling with ringing artefacts, especially with the overlaid diagram thingies in the first few frames.

i can post a chunk if you like :)

[ObiKenobi]

Originally posted by Mug Funky
@ obiwan: look harder :)

I looked real hard. There was none.

Originally posted by Mug Funky
most notably the intro sequence - the title overlays are 30i interlaced in places (can you believe that?)

Not surprised at all. That's how most of the overlays are done.

Originally posted by Mug Funky
i can post a chunk if you like :)

Please do.

[Chainmax]

It's a damn shame to hear that the US releases are so screwed up, from what I heard the R2 releases are completely progressive.