[MPlayer-users] Re: Crop before deinterlacing OK?

[Rich Felker]

On Mon, Jul 25, 2005 at 12:30:11PM +0200, Matthias Wieser wrote:
> > Just now I tried pp=lb vs. kerndeint and found pp=lb to produce 45%
> > smaller files (using vqscale=2) in a high-ghosting clip.
> 
> During high motion scenes kerndeint has pp=ci characteristics. So it 
> throws away halve of the lines and does an interpolation. So there are 
> interpolation artifacts and there is no noise-smoothing by blendung two 
> fields into each other. Additinally the local contrast is not lowered 
> (think of a black, 1 pixel high horizontal line on white ground which 
> becomes gray and blurry after applying pp=lb)
> 
> => no question, the output of kerndeint during high motion scenes is more 
> noisy, has higher contrasts and some interpolation artifacts which all 
> hurts compressibility. But on the other hand you have the advantage of 
> nearly unaltered frames during no motion scenes.
> 
> Maybe kerndeint could be improved by replacing the cubic interpolation by 
> the pp=fd interpolation which gives nicer edges.

pp=fd is not (just) interpolation. It's a blend filter, just like lb,
except it only operates on one field and blends the interpolated value
with the discrete laplacian of the other field rather than the
original pixels of the other field. This makes nasty ghosts, in the
form of an edge-detected version of the discarded field...

> Maybe someone ports some additional deinterlacers from dscaler...

The only way to get anything remotely correct is by doing per-pixel
motion estimation using 2 fields of same parity, and then interpolated
motion compensation on the opposite parity field to try to make it
line up.

Rich