Why is rasterization inconsistent and unbiased? If we can take infinite number of samples, then we can basically recover the original image right?
@pavelkang: No matter how many samples we take, we will never get light paths like light bouncing off one wall onto another, or light focusing through glass to make a caustic. It is the fact that we miss these important paths that makes the image unbiased. (To be fair, I should probably say that "rasterization" here really means "rasterization using the standard lighting model of the OpenGL pipeline," i.e., local lighting computed from the normal, eye, and light direction. One could in principle write a physically-based renderer using rasterization simply as a tool.)
Is pathtracing "almost" only because of lights that are deltas and materials that are deltas? If we use a technique to get around that, isn't the pathtracer consistent and unbiased?
In assignment 3, since we shoot rays at delta lights and sample delta materials carefully, is our pathtracer consistent and unbiased? It is my understanding that it is.
Completely understanding sources of bias in Monte Carlo rendering can be a rather tricky issue. For instance, suppose you use interpolated vertex normals for shading---already your result will be biased, because the illumination does not correspond to the actual geometry. So, you have to be careful. If you want to understand all the issues in great detail, a great place to look is the thesis by Veach.