[Theora-dev] Re: [Fwd: Re: Theora mcomp tuning...]

[Timothy B. Terriberry]

Adam D. Moss wrote:
>> Although SSD may seem to be a nicer theoretical quantity to work with, 
>> this is only true when errors come from a Gaussian source.
> 
> 
> Okay -- I guess that's true of most of my experiments on estimation
> of perceptible quantization error.

Motion compensation does not deal with quantization error, but 
prediction error, which has altogether different statistics. The errors 
are better approximated by an exponential distribution, for which SAD is 
the optimal metric. It's still only an approximation. A Cauchy 
distribution can do an even better job, but the optimal error 
metric---log(1+d^2/a^2) for some constant a---imposes too heavy a 
computational burden to be practical, and the gains are minor.