[FFmpeg-devel] [PATCH] Apple RPZA encoder

[Kostya]

On Thu, Apr 16, 2009 at 04:21:39PM +0530, Jai Menon wrote:
> On 3/30/09, Michael Niedermayer <michaelni at gmx.at> wrote:
> > On Mon, Mar 30, 2009 at 10:17:43AM +0530, Jai Menon wrote:
> >  > On 3/25/09, Michael Niedermayer <michaelni at gmx.at> wrote:
> >  > > On Tue, Mar 24, 2009 at 03:39:11PM +0530, Jai Menon wrote:
> >  > >  > On 3/24/09, Jai Menon <jmenon86 at gmail.com> wrote:
> >
> > [...]
> >  > >  > +/* tuning parameters */
> >  > >  > +#define SIXTEEN_COLOR_THRESH        24
> >  > >  > +#define SKIP_FRAME_THRESH           13
> >  > >  > +#define START_ONE_COLOR_THRESH      8
> >  > >  > +#define CONTINUE_ONE_COLOR_THRESH   0
> >  > >
> >  > >  Encoders, especially such simple ones should be build to be RD optimal
> >  > >  in as many aspects as possible
> >  >
> >  > The "parameters" so to say, give a good tradeoff between final output
> >  > size and quality. But I don't think I understand correctly. If we were
> >  > to use 16 colors everywhere, the output size will be absurdly large.
> >
> >
> > please read doc/rate-distortion.txt
> >  and if you dont understand it please try to explain what you dont understand
> >  so i can improve it
> 
> I took some time off to wade through that doc. From what I understand,
> I'll need a metric to represent the distortion and output rate.
> Assuming I use a simple variant like least squares or sum of absolute
> differences or whatever, I still don't understand how I can visualize
> and RD curve for this (rpza) case. Do I fix the output size and
> iterate over possible values for the lagrangian parameter to minimize
> the bit cost? If I do that, I still don't quite get how I can
> correlate that bit cost to rpza block coding. Any insights appreciated
> :)

You have two parameters - bit cost and distortion. Bit cost is number of bits
needed to code this block with method X (fill, skip, N colours) and distortion
may be squared error between original and coded block. I had something like
that for MSVC encoder.
 
> -- 
> Regards,
> 
> Jai