[MEncoder-users] High quality encoding of anime and broken mencoder options

[Loren Merritt]

On Wed, 5 Mar 2008, Rich Felker wrote:
> On Wed, Mar 05, 2008 at 12:37:29PM -0300, Jorge Peixoto de Morais Neto wrote:

>>>> Hi. I seek high quality encoding of anime. I have tested
>>>> many encoding options, always with vcodec=mpeg4 and 2-pass encoding.
>>>> I have found that even with vbitrate, the actual bitrate depends on
>>>> the options used. Therefore, to keep the comparison fair, I have
>>>> measured encoding quality by Q=PSNR/bitrate, with bitrate in Mbits/s.
>>>
>>> I believe that your Q value is not relevant: why PSNR/bitrate rather than
>>> PSNR²/bitrate or PSNR/bitrate², or PSNR/log(bitrate) or... PSNR itself is a
>>> logarithm of mean square error, there is no reason that a simple division
>>> should be the relevant measure.
>>
>> The Shannon-Hartley theorem
>> (http://en.wikipedia.org/wiki/Shannon-Hartley_theorem) suggests to me
>> that information is proportional to the PSNR (I know I am stretching
>> things here, and I may be wrong).
>
> Amount of information and quality are not necessarily very close when
> it comes to human perception.

You don't even have to get into psychovisual arguments.
PSNR/bitrate is wrong because a _factor_ of X in bitrate corresponds to a 
_difference_ of Y in PSNR, for some content- and codec-dependent values of 
X and Y. And you don't need theory to prove this, just encode something at 
multiple bitrates.

--Loren Merritt