[MPlayer-users] Re: [-] Best de-interlace filter ( NOVIRUS )

[D Richard Felker III]

On Tue, Oct 12, 2004 at 03:06:40PM +0000, james168 wrote:
> > On Mon, Oct 11, 2004 at 09:31:48PM -0700, Jack wrote:
> > > Why no mention of lavcdint... does it just suck?
> > 
> > yes.
> 
> What's wrong with it?

there's absolutely no logical basis for what it does. the basic
principle goes like this: it uses the second derivative of the field
being thrown away, along with interpolation of the field being kept,
to make the new replacement field. in the case where there's no
motion, this will avoid losing all the information from the field
thrown away so it doesn't look so bad (but still not as good as no
deinterlacing at all in this area, or linear-blend deinterlacing in
this area). on the other hand, if there is motion, there's absolutely
no reason for there to be a correspondence between the second
derivative of one field and the missing lines in the other field.
using this data is totally bogus, and although it won't be terribly
visible with blurry content, it will create horrible ghost outlines
wherever the thrown-away field had sharp contrasted borders. even if
these artifacts aren't very visible to your eye they'll waste lots of
bits to encode.

if you want to see what i mean very clearly, try lavcdeint (better
known as pp=fd, the fast version of the identical algorithm) on
telecined anime and watch what happens to the fine outlines on
interlaced frames... the same is happening everywhere, although
often to a lesser extent.

anyway, furthermore lavcdeint is a deinterlacer that throws lots of
information away. it basically gets rid of one field, except it keeps
the second derivative from it to predict the missing lines of the
other field, which i explained is wrong in general. so it's bad like
li or ci for throwing things away, _and_ bad like lb for creating
ghosts... it's just bad in every way a deinterlacer can be bad.

rich