[XviD-devel] Re: mrSAD

[Radoslaw Czyz]

Hi

gruel wrote:

>> Why are you so sure mrSAD is better?

I'm not. ;-)

>>  After all, you throw away some information (absolute lumi) when using it.

As you might remember, I've created mrSADed pmvfast already. I wasn't
throwing away any information - I was adding absolute lumi to SAD,
multiplied by a constant.

>> I am very sure that in scenes with changes in brightness mrSAD is
>> better than SAD, but in general I would not say so.

That's very possible. However, my first mrsad gave PSNR improvement
for all scenes. I might be wrong, but I have a feeling that mrSAD
tries to match a texture, while SAD tries to match brightness. We want
texture, of course.

>> Also, I guess you read that mrSAD has a more difficult error landscape
>> with many more local minima. How are you going to avoid those? The
>> hierarchical subsampling idea? 

I'm sure that mrSAD will be worse for noisy data.
In a perfect world, we would make SAD-ed ME first, and then use the
result as a prediction for mrSAD-ed ME. This is a theory because of
computing power.

I'm not saying that mrSAD is better, but I'm going to give it a try.
As you can see, I've created not-mrSADed motion procedure first, so
that I can compare mrSAD and SAD in the same procedure. If I'm wrong
about mrSAD, I won't cry.


Radek