For those who don't work in graphics, a 'noise' function is a pseudo-random continuous, and usually differentiable, function. The most fanous example is Perlin noise. Typically a noise function is first defined by data defined on a lattice in R^n, eg. through some kind of hash function (ie. a function whose value is 'hard' to predict) and then extended to all of R^n via some kind of filter - for example fitting a smooth Catmull-Rom spline through the data points. These types of function are ubiquitous in computer graphics where they can be used to synthesise anything from the shape of a simulated tornado to the distribution of dirt on a car body.

Unfortunately one of the problems with most noise functions is that they are hard to 'anti-alias'. When you render a 3D image you typically need to colour an individual pixel by integrating over all of the light that falls onto it. The light that falls on a single pixel is typically the integral of the colour of an object over a small part of its surface - and this colour is often given by a noise function of some sort. The problem is that integration is expensive and so people often just 'fire' a bunch of rays from the pixel into the object and average the colour of the ray-object intersections. But we know from the Nyquist

theorem that if you don't sample the colour enough you may end up getting high frequency garbage 'aliasing' as a low frequency signal. With regular colour patterns like checkerboards you get moire patterns but with noise functions you get unwanted low frequency noise. What we really need is a noise function that is band limited so that the high frequency components simply aren't present to cause aliasing. Perlin noise isn't band limited so that if you try to reduce the amount of it to suppress aliasing you end up not having any pseudo-random detail at all. In Wavelet Noise the authors explicitly construct a noise function that has a hard frequency cutoff. It becomes a true band limited function and hence we can get a reasonable approximation to its integral by sampling.

The great thing about this paper was its presentation. It had a plot. As the publication deadline approached the authors realises their noise function had a flaw. Typically we're not interested in sampling the value of the noise function in R^3 but over an R^2 surface embedded in R^3. The catch is that such a 'slice' of a band limited 3D function is not a 2D band-limited function. This meant that their noise function couldn't actually be used to colour the surfaces of objects! Fortunately they figured out the solution: instead of colouring the surface simply by pulling back the colour from the value of noise in R^3, you instead integrate along the normal to the R^2 surface. The Fourier transform of such an integral is simply a slice through the Fourier transform of the original function on R^3 and hence is band limited. In the nick of time the authors figured this out and then amazingly the authors were able to implement this with a small piece of code that runs reasonably fast.

By the way, this is also where the paper on Dual Photography was presented. I mentioned that a few months ago as I copied one of their results here.

Apologies: this entry probably has many typos and other kinds of error. For some reason the Safari web browser refuses to work at my hotel and for some reason I'm having trouble editing text in Firefox.

## No comments:

Post a Comment