pullup: Cannot find a way to say “bottom field first”

[ion1]

I was unable to find a way to play frames_of_reference.mpeg in mpv without telecine artifacts. --field-dominance does not seem to affect pullup. None of the parameters to pullup seem to help either. autodeint.lua was also unable to come up with a working filter.

I was able to get the video corrected to various extents with ffmpeg:

• detelecine=first_field=b:start_frame=1 would remove all artifacts until going out of sync towards the end of the video.
• fieldmatch=order=bff,decimate would remove most artifacts throughout the video, leaving an interlaced frame here and there.
• fieldmatch=order=bff:combmatch=full,yadif=deint=interlaced,decimate would fix many of those frames but not all.

I suppose mpv pullup with a field dominance override would be more or less equivalent to the second filter? I would be happy with that.

I’m running mpv git-eabc530.

[richardpl]

use setfield filter before pullup

[ion1]

Thanks! For future reference, I got good results (similar to the ffmpeg output) with

mpv --vf=lavfi=\"setfield=bff,pullup,dejudder\"

Perhaps mpv could provide a higher-level setting that is documented in the man page. But if you deem this use case too obscure, feel free to just close the request.

[kevmitch]

so --field-dominance=bottom doesn't work for pullup?

[kevmitch]

oh, duh, yeah, that's what you said. yeah, this should work, looking into it now.

[kevmitch]

Well the only place that vf_pullup.c references the field dominance is
in filter_frame(AVFilterLink *inlink, AVFrame *in) and that is definitely influenced by setting --field-dominance in mpv.

It looks like some of the supposedly progressive frames in the 2-3 pattern are actually interlaced. So even if pullup could perfectly figure what to make of it, you'd still get interlaced frames. You're probably better off just using yadif on it.

[ion1]

With ffmpeg, detelecine=first_field=b:start_frame=1 (which to my knowledge is not adaptive but blindly goes through the motions of inverse telecine) seemed to remove all the artifacts until going out of sync minutes later. This would seem to indicate a pure 2-3 pattern.

The adaptive filters (fieldmatch+decimate and pullup) seemed to fail occasionally, but had way better results than yadif. The failure frames were infrequent enough not to ruin the experience.

Both --vf=lavfi=\"setfield=bff,pullup\" and --vf=lavfi=\"setfield=bff\",pullup worked but I also had to add dejudder to make the motion smooth.

(Is pullup ever better than pullup,dejudder? Should the mpv pullup filter perhaps activate both unconditionally?)

[kevmitch]

It looks like one of the problems with that clip is that all the frames are flagged progressive, so the field dominance is actually being ignored in vf_pullup.c even though it is being correctly set by --field-dominance in mpv. The parity calculation in vf_pullup.c is

p = in->interlaced_frame ? !in->top_field_first : 0;

The reason that --vf=lavfi=[setfield=bff,pullup] works better is that setfield has the side effect of also unconditionally flagging all frames as interlaced so that the field dominance will actually be honoured. @wm4 should the --field-dominance option do the same, should there be a separate option to reflag frames as interlaced/progressive, or should we just continue the status quo where this has to be done in lavfi somehow?

That said, I still find that yadif is at least required at the end to make that clip watchable --vf=lavfi=[setfield=bff,pullup,dejudder,yadif] due to the significant residual combing.

[ghost]

I wonder about that. Just using setfield seems like a decent solution too?