Why don't the built-in resizers use a 2D kernel?

[mimungr]

I'm digging through the AVISynth source code. If I'm reading it correctly, the built-in resize filters don't do a 2D resize. Instead, they resize horizontal and vertical independently. Is this purely for performance reasons? I would think that a 2D kernel would yield better results. I'm tempted to implement such a filter, but don't want to waste my time if that avenue has already been explored and rejected. I'm particularly interested in a 2D version of Lanczos3.

[Manao]

The kernel used are separable, so the only loss in quality is due to rounding to a 8 bits integer between the horizontal and the vertical step.

[mimungr]

I understand that. But wouldn't a non-separable 2D kernel yield better results?

In other words, AVS is effectively using f(x) + f(y). I'm proposing f(sqrt(x^2 + y^2)). So the Lanczos window is a circle, not a square, and the matrix isn't separable.

Does AVS use a separable kernel because it yields better image quality? Or because it's good enough and a non-separable kernel is too slow?

[Manao]

mimungr : it's not f(x) + f(y), it's f(x)*g(y) ( else it wouldn't be separable ).

Anyway, a separable kernel is chosen because pixels are squares, not circles, and the sampling grid is made of squares. ( that's the quickest explanation without having to go into the fourier space and having to write some formulas )

[Fizick]

I always wondered too, why do not exist some simple non-separable 2D resizer filter (similar to bi-linear for example). Or may be it is named "precise" bilinear?

[MfA]

Anyway, a separable kernel is chosen because pixels are squares, not circles, and the sampling grid is made of squares. ( that's the quickest explanation without having to go into the fourier space and having to write some formulas )I dunno ... I think if you implemented something like lanczos with a circularly symmetric sinc function it would look a little better, just not enough better to justify the cost.

[IanB]

@mimungr,

If you are so keen knock yourself out. However with the use of only 8 bit data the difference will struggle to exceed the 1/256th needed to be detectable and it will probably be as slow as ...

If 15 bit data is ever implemented you might have a case, but you are probably going to want to start with 5 or 6 tap filters first.