[MPlayer-users] AVI encoded file playback on other systems

[D Richard Felker III]

On Wed, Sep 24, 2003 at 07:29:14PM +0200, Karl Ewald wrote:
> [Automatic answer: RTFM (read DOCS, FAQ), also read DOCS/bugreports.html]
> Corey wrote:
> >Actually, you really should be using 2-pass mode to encode anything you
> >want to keep. On the first pass, the ratecontrol algorithms can only
> >guess which future parts of the video will need a higher bitrate (such
> >as when the picture changes rapidly) and which will need only need a
> >small amount. The second pass then uses statistics from the first pass
> >to better allocate data, resulting in a file with more consistant
> >quality.
> 
> As far as I understand, if I use constant quantizer I should get uniform
> quality at the expense of unpredictable filesize. So it allocates the amount
> of data needed for the requested quality anyway, and two pass mode wouldn't
> make any difference. Only if I want to match a target filesize, two-pass
> would raise or lower the quality uniformly to make the resulting file match
> the target size.
> 
> While I do not specify a constant quantiser with vqscale, I use very narrow
> bounds for the quantiser, e.g. vqmin=7:vqmax=8. With a (too) small target
> bitrate this should act identical to vqscale=8 since the vqmin/vqmax
> corridor takes precedence over the target bitrate, and my remark above
> applies.
> 
> I can live with the variable filesize resulting from this approach, since
> for given material  (e.g. episodes of the same series) I quickly gather an
> idea of the range of filesizes usually produced.
> 
> Is what I'm writing here basically correct or am I mistaken?

Using qmin/max like this will give much worse quality than doing 2pass
rate control the right way, even if both result in the same size. This
is because q=7 may look very bad during stills while q=8 may be
overkill for encoding frames with very rapid motion. Since q=2 or 3
doesn't take that much space during simple low-motion scenes, it makes
sense to let the ratecontrol code use low quantizers (at least down to
3, or preferably 2) when it wants to.

Rich