Better alternative for repairing badly upscaled chroma on interlaced NTSC DVD?

[Reel.Deel]

Currently working on some of my DVDs and I came across one where the chroma has been upscaled horribly (PointResize? at least that's what it looks like). This is a 1st for me, even though I have something that works decently (little blurry) I was wondering if you guys have better alternatives or suggestions? Here's a brief sample (https://dl.dropbox.com/s/9ifarh3jwbyxbwj/sample.m2v) showing the problem.

This is my current script, not really sure why it works :p but here's the output before (https://dl.dropbox.com/s/lgyw46notbqfhb0/Before.png) and after (https://dl.dropboxusercontent.com/s/9vmgb56afaehbom/After.png).
MPEG2Source("sample.d2v")
SeparateFields()
o = last

ConvertToYUY2(chromaresample="Point")
ReInterpolate411()

U = UToY8().BicubicResize(width/2, height/2, src_left=-0.50, src_top=1.0)
V = VToY8().BicubicResize(width/2, height/2, src_left=-0.50, src_top=1.5)

YToUV(U, V, O)
#Weave()
#QTGMC()

#debug only
Interleave(o,last)
ConvertToRGB32(chromaresample="Point")
PointResize(width*2, height*2)

[feisty2]

pointresize is reversible, if you're sure it's upscaled with pointresize, you can pointresize it back

[Reel.Deel]

It looks like PointResize but may have had other processing done to it so it's not completely reversible. Regardless, I already tried that and the results were not as good as my current script.

[feisty2]

little bit confused here, your clip is an mpeg2 video, so I guess it would be yv12 in the most possible way, and how could yv12 chroma be upscaled from some original video, shouldn't it be downscaled from yuy2 or yv24?
edit: is your clip upscaled?

[Reel.Deel]

Well that's assuming that this DVD came from a perfect world, I think we all know that's not the case. Proprietary software/hardware and industry 'professionals' find a way to mess things up. This footage was shot circa 2003 (small budget) so if I had to guess I would say that it was originally DV/DVCPRO 4:1:1 and then blindly converted to 4:2:2 for editing and finally 4:2:0 for DVD. Don't know how correct I would be but problems with DV and wrong chroma resampling was a known problem back in the day. That's why plugins like ReInterpolate411 (http://forum.doom9.org/showthread.php?p=352001#post352001) and ReInterpolate420 (http://forum.doom9.org/showthread.php?t=82787) exist.

[feisty2]

I tried to pointresize chroma back and upscale it with spline64resize, it works perfectly, so I don't know what you mean pointresize doesn't work

separatefields ()
y=grayscale ()
u=utoy ().pointresize (180,120).spline64resize (360,120)
v=vtoy ().pointresize (180,120).spline64resize (360,120)
ytouv (u,v,y)
weave ()

[Reel.Deel]

I never said PointResize didn't work, I said the results weren't as good. Also, I tried something a little differently, I downscaled with PointResize and then upscale chroma with nnedi3_rpow2 but the chroma was bobbing (don't know why). Anyways, I guess sometimes it's best to keep things simple. You're script does a pretty good job. Do you by chance know the offsets for the chroma placement to be MPEG2?

[feisty2]

it's a nnedi3 issue, you can't use nnedi3_rpow2 to upscale chroma since it's a top left aligned resizer (the align is designed for mvtools2 or awarp4 kind of supersample interpolation, not general image enlargement), it will get you center shift because you need a center aligned resizer, you gotta adjust the nnedi3 result with other resizers to fix the shift, like spline64resize (width,height,src_left=-0.25,src_top=-0.25)

[colours]

If the chroma planes end up being the same resolution as the source, there's no potential chroma shift to care about from ignoring MPEG-2 chroma siting.

For that matter, using PointResize to downscale only to upscale again with nnedi3_rpow2 is wasteful: just use nnedi3 directly!

ffvideosource("O:\sample.m2v")
nnedi3(field=3) # deinterlace first
u = utoy().turnleft().nnedi3(field=0).turnright().spline36resize(360,240,src_left=-0.5) # antialias the chroma
v = vtoy().turnleft().nnedi3(field=0).turnright().spline36resize(360,240,src_left=-0.5)
ytouv(u,v,last)
qtgmc(inputtype=1) # then use qtgmc in progressive mode

[feisty2]

that's no good, most people will not use pure nnedi3, one single intra filter to deinterlace for quality concerns, they prefer to use complex script for INTERPOLATION, not just STABILIZATION

[colours]

So you're saying that most people use QTGMC(Preset="Ultra Fast") because that's the only QTGMC preset that uses a not-purely-spatial deinterlacer for the first step?

I find that a bit hard to believe!

[Reel.Deel]

...using PointResize to downscale only to upscale again with nnedi3_rpow2 is wasteful: just use nnedi3 directly!

Thanks for the suggestion, believe me or not this is very close to what I wanted to do but had trouble making it work. I was more under the impression that using the progressive modes on QTGMC was more geared towards badly deinterlaced content. So is QTGMC(inputtype=1) the same as QTGMC() minus nnedi3 or is there more missing?

PS: Loving the signature :)

[feisty2]

the interpolation is not pure nnedi3/eedi3, it's actually edi+mflowinter, that's why it's called qtgMC, I think I read the qtgmc script before, now I can't remember all those specific details, but I guess it approximately works that way
edit: I'll check qtgmc script later when I get to my computer, and see what exactly does qtgmc do if inputtype=1

[feisty2]

inputtype=1 is designed to STABILIZE low quality deinterlaced clips, I guess it's not the same as normal interpolation process

[feisty2]

okay, I just checked qtgmc script again, it doesn't use mflowinter, did a little search on forum, mcbob uses it instead, seems like I mixed them up in my memory, and yes, you can use inputtype=1, it's identical to qtgmc internal process, tho I guess nnedi3 is not a good deinterlace kernel since it's too weak at connecting lines which is one of the most important features required for a deinterlace kernel, I always thought nnedi3 is more like an upscale kernel, back to the point, eedi3 (sclip=nnedi3 ()) would be a lot better for deinterlacing, and dude, you don't really wanna upscale chroma with nnedi3 cuz chroma quality is usually..., well, we all know the answer, I think a smooth algorithm like eedi3 would be better for chroma upscaling