[Ffmpeg-devel] Trellis Quantization Applied To (A)DPCM

[Michael Niedermayer]

Hi

On Tue, Jun 06, 2006 at 04:09:38PM -0700, Loren Merritt wrote:
> On Mon, 5 Jun 2006, Michael Niedermayer wrote:
> >On Sun, Jun 04, 2006 at 10:01:01PM -0700, Loren Merritt wrote:
> >>
> >>Added some asymmetric tables. They helped only a little, but I don't
> >>really know what to optimize for when writing a table by hand.
> >>
> >>OTOH, a brute-force iterative search isn't too ridiculously slow.
> >>The next step would be to run a clustering algorithm like I proposed 
> >>before, but this time looking for N tables to jointly fit all the
> >>frames of a training set, instead of clustering based on the statistics 
> >>of one frame at a time.
> >
> >the iterative search seems to get stuck in local minima ...
> >[...]
> 
> And now with the clustering.
> 
> exponential tables
> context=0 psnr:43.42 tables_tried:0
> context=1 psnr:43.56 tables_tried:3326
> context=2 psnr:44.43 tables_tried:57057
> context=3 psnr:45.85 tables_tried:399119
> 
> clustered tables
> context=0 psnr:46.20 tables_tried:0
> context=1 psnr:47.11 tables_tried:2928
> context=2 psnr:47.34 tables_tried:82576
> context=3 psnr:47.45 tables_tried:414024

impressive

btw, if i outcomment the following then i see a 0.21 PSNR gain for
context=2 with my test file ...
    if(v>=-2 && v<=2)
        continue;

[...]

-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

In the past you could go to a library and read, borrow or copy any book
Today you'd get arrested for mere telling someone where the library is