Filtering and Sampling (Comp238) :

In this assignment I've investigated a few filtering and sampling techniques. I've used the pbrt code as a start. The table below shows the result of applying one of the filters - box, gaussian, mitchell, window sinc and lanczos Sinc, with one of the sampling techniques uniform, random and halton. A window sinc filter differs from the Lanczos sinc in that it uses a hamming window in place of the lanczos window.

Each column represents one of the sampling techniques while each row represents one of the reconstruction techniques. The samplers use 4 image samples either 4 across the whole pixel or 2 in each dimension.

   uniform random halton
 box  Filter:box Sampler:uniform Filter:box Sampler:random Filter:box Sampler:halton
 gaussian  Filter:gaussian Sampler:uniform Filter:gaussian Sampler:random Filter:gaussian Sampler:halton
 mithcell  Filter:mithcell Sampler:uniform Filter:mithcell Sampler:random Filter:mithcell Sampler:halton
 windowsinc  Filter:windowsinc Sampler:uniform Filter:windowsinc Sampler:random Filter:windowsinc Sampler:halton
 lanczossinc  Filter:lanczossinc Sampler:uniform Filter:lanczossinc Sampler:random Filter:lanczossinc Sampler:halton

A close to ideal image using the mitchell reconstruction filter and the halton sampler with 64 samples per point looks like ...
Ideal..

The performance difference between the two filters windowsinc and lanczossinc isn't really much. They both have some ringing effects towards the horizon. The mitchell filter on the other hand has much less ringing effects. The gaussian reconstruction filter seems to be avoiding the artifacts by blurring out several details near the horizon.

The uniform sampler is clearly bad (esp alongwith the box reconstruction filter) It results in a lot of "patchy" kind of rendering towards the horizon. The random sampler performs better than the uniform one, but the noise added can be easily perceived especially towards the horizon. The noise is however a little less pronounced in conjunction with the mitchell filter.

Due to the artifacts near the horizon, I felt that the sinc filters didnt perform real well, whereas the halton sampler in conjunction with the mitchell filter performed relatively better.

As for the 64 sample case, the combination of the halton sampler and the mitchell filter worked real well almost all the way till the horizon. The frequencies near the horizon are very high. The halton sampler is quasi-random and the mitchell filter has a better fall-off but it still allows for some ringing. But still the artifacts dont seem to look like ringing, my guess would be that they have something to do with computation.

Rohit Saboo
Generated on linux running on an Intel Xeon 2GHz machine (swan)
Uses source code from Physically Based Rendering, Pharr and Humphreys