What is this type of defect called and how to get rid of it?

[takla]

Sample here

I'd like to get rid of these vertical lines that go through the whole image, especially visible on the painting on the wall in the back, during motion.

I'm pretty sure thats not just aliasing. If you pay attention, you can see those lines go across the whole frame.

[Emulgator]

Aliasing I guess still.
Non-matching horizontal scanning sensor resolution -> in-cam application of any inadequate simplesimon kernel to obtain -> any 704/720 horizontal SD resolution.
Then uprezzing SD -> HD and making it worse.
Solution: Find all conversion parameters, roll back reversing every step in backwards sequence and go from there.

[takla]

BicubicResizeMT(768, 1080, b=1, c=0).Spline36ResizeMT(1920, 1080)

This works o.k.
But I can't get this low quality source up to my standards.
Really sucks that all good media has mostly garbage sources.
Oh well...

[poisondeathray]

Another approach is to use some sort of temporal smoothing, at the risk of detail loss

Normally you could use QTGMC in progressive mode in avisynth, but it doesn't do well here without causing significant damage. I tried turning it left/right (to make the lines horizontal) , also downsizing

vs_temporalfix for vapoursynth does a decent job on this scene without turning it to mush . Test it on other sections and adjust the settings appropriately.
https://github.com/pifroggi/vs_temporalfix

https://www.mediafire.com/file/vu8rg...0_tr6.mkv/file

[takla]

Originally Posted by poisondeathray

vs_temporalfix for vapoursynth does a decent job on this scene without turning it to mush . Test it on other sections and adjust the settings appropriately.

Amazing. Thank you very much.

[takla]

vs_temporalfix just got an update and can now work with the GPU. simply amazing.

[geometer]

This is great, it defeats the temporal instability that comes from shapes toggling between left and right pixel position in a coarse grid.

But I suppose you want to reduce the ringing also, ..
now I'm not prepared for this type of TV signal, but you will see the point:



FFmpegSource2("sample (1).mp4")

last = AntiRingLR2(last, "luma", weightL=1.8,offsetL=4,weightR=1.5,offsetR=5,knee=0.15,pr2=0.1,pr3=0.7, output="out=in")

last = AntiRingLR2(last, "luma", weightL=1.3,offsetL=0,weightR=0.9,offsetR=0,knee=0.1,pr2=0.1,pr3=0.1, output="out=in")

return last


It will touch the issue, but needs complete processing of tails, this time on the left side, while I have coded so far only the right side tail.
This is a matter of time for another small coding session, anyone can do it.
But fine-tuning takes some experience with this kind of tool.
You will probably see some improvement, but not enough, still shows proof of concept.

One of the problems to apply it here is that it cannot derive a precise model of the edge (its slant etc), after the fact, after the upsizing, which makes coarse guesses only.

This example (original posting) in my view has 3 issues:
1. ringing (as a net sum of many filtering and encoding processes)
2. some sort of moiree because of switching the sampling domains, upsampling etc., and of course from hatched shapes that the video cam tried to fit into its grid.
3. the typical wiggling and bobbing for which vs_temporalfix is an optimal cure.

What I suggest is pre-processing with some variant of this AntiRingLR() that I have introduced in the other Expr() tool thread.
( https://forum.doom9.org/showthread.php?t=186925 )

[tormento]

It would be really nice to have a porting of vs_temporalfix for AVS+ to test.

Thanks to anybody who will provide.