Re: MMX speedup for Floyd Steinberg error diffusion

rep_movsd  <spamtrap@[EMAIL PROTECTED]
> writes:
> On May 10, 10:48 pm, Phil Carmody <thefatphil_demun...@[EMAIL PROTECTED]
>
> wrote:
>>
>> You might be able to just do some kind of octree algorithm.
>> You could even hard-code the partitions.
>>
>> (Octree's also possible in 256-colours too, of course. You can
>> even create a slightly optimised palate by using 2 p***** and
>> crudely bin-counting on the 3 axes.)
>>
>
> I implemented 4 different quantizers, Octree, Median cut, Wu and
> NeQuant ( the last 2 adapted from the FreeImage library ) and it seems
> that Wu quantization gives the best results, followed by octrees, The
> other 2 never quite get as good results as far as I could see.
>
> What do you mean by "bin-counting on the 3 axes" ?

Bin counting is just building a histogram. So pretend 
there's complete independence on the 3 axes, and 
populate 3 histograms with tallies. Decide how to split
on each axis based purely on its own histogram. 
This works quite poorly if there's a very high correlation
between the three axes. Mapping to some kind of 
luma-chroma space first can help fix that.

Phil
-- 
Dear aunt, let's set so double the killer delete select all.
-- Microsoft voice recognition live demonstration