TABLE OF CONTENTS The basic algorithms in the previous sections can be made more effective by using variance reduction techniques and low-discrepancy sampling. In this section, we will discuss variance reduction by view-importance sampling, by control variates, by combining gathering and shooting estimators using the same random walks or rays, and by weighted importance sampling. The material covered in this section is of a great practical importance and also serves as an illustration of the variance reduction techniques discussed in Chapter 3.
In the basic algorithms in the previous sections, transitions are sampled using probabilities that reflect the laws of physics. The quality of the computed result mainly depends on the area and reflectivity of the patches but is furthermore uniform in the whole scene. Sometimes, however, one would like to save computation time by having high quality only in a part of the scene, for instance, the part of the scene that is visible in a view, while compromising on the quality in unimportant parts of the scene (see Figure 6.24). For example, when computing an image inside a single room in a large building with several floors, each containing many rooms, the basic estimators would spend a lot of work in computing the illumination in all rooms on all floors to similar quality. One might prefer to concentrate the computation work on the room one is in, at the expense of a lower quality of the radiosity solution in other...
