Re: transcode filters

On Wed, Apr 21, 2004 at 08:29:31PM +0400, Vladimir Mosgalin wrote:
> On Wed, 21 Apr 2004, D Richard Felker III wrote:
> 
> DRFI>Hmm, but you shouldn't need a complicated filter chain like that to
> DRFI>remove telecine. The steps make no sense...
> 
> Nope, it's just transcode way of performing ivtc. It is written in
> documentation
> ---
>    Additionally, when the telecine pattern shifts (as a result
> of, maybe, field editing by the DVD authors) a couple of
> interlaced frames will pass through. These should be deinterlaced 
> on their own. You shouldn't use global deinterlacing after ivtc 
> - it would blur out the otherwise perfect progressive frames 
> that get reconstructed from 'ivtc'.

This most likely indicates that the ivtc filter is buggy. It should
never output any interlaced material for purely telecined content. If
the original is a mix of telecined film and video, then all bets are
off...

> However, transcode's '32detect'
> filter comes to the rescue: It first checks whether the frame 
> is interlaced, and only then it forces a frame deinterlacing. 

This will result in ugly flicker in your movie where frames
incorrectly get deinterlaced rather than properly matched. You'd be
better off deinterlacing all the frames with a blend filter so it's
uniform at least...

> So, this is what I recommend for your sessions:
> 
> transcode -M 0 -f 23.976 -J ivtc,32detect=force_mode=3,decimate ...

I still don't understand how decimate is useful...

> DRFI>Ah. li and ci are correct deinterlacers but the output looks "low
> DRFI>resolution". lb is incorrect but looks ok if there's only low motion.
> DRFI>fd (aka ffmpeg deinterlacer, aka lavcdeint, which kerndeint seems to
> DRFI>mimic) is completely wrong (but based on formulae in some mpeg
> DRFI>spec...) and produces a ghost that looks like an edge-detect image of
> DRFI>the other field over top of the picture.
> 
> Well, result from kerndeint looks almost like cubic interpolation to me.

It's definitely not. As far as I can tell it's the same
lowpass/highpass thing the Decomb author described on that web site
you cited. Cubic would "look bad" but not create artifacts.

> I liked smartdeint in blend mode more...

I agree. The kernel they use is absolutely horrible, but lots of
people seem to like it for some dumb reason. I think it's supposed to
look ok with very-low-motion live action video, but it's very very bad
on animation or high-motion video.

Rich