Analyse spatiale et spectrale des motifs d'échantillonnage pour l'intégration Monte Carlo. (Spatial and spectral analysis of sampling patterns for Monte Carlo integration)

Sampling is a key step in rendering pipeline. It allows the integration of light arriving to a point of the scene in order to calculate its color. Monte Carlo integration is generally the most used method to approximate that integral by choosing a finite number of samples. Reducing the bias and the variance of Monte Carlo integration has become one of the most important issues in realistic rendering. The solutions found are based on smartly positioning the samples points in a way that maximizes the uniformity of the distribution while avoiding the regularities. From this point of view, the 80s were a turning point in this domain, as new stochastic methods appeared. With a better comprehension of links between Monte Carlo integration and sampling, these methods allow the reduction of noise and of variance in rendered images. In parallel, the complexity of sampling methods has considerably enhanced, enabling to have fast as well as good quality methods. However, these improvements have been done by trial and error focusing on two major points : the improvement of sampling pattern uniformity, and the suppression of regularities. Even though there exists some theories allowing to bound the error of the integration, they are usually limited, and even inapplicable in computer graphics. This thesis proposes to gather the analysis tools of sampling patterns and to connect them together. These tools can characterize spatial properties such as the distribution of distances between points, as well as spectral properties via Fourier transformation. Secondly, we have used these tools in order to give a simple expression of the bias and the variance for Monte Carlo integration ; this is done by using prerequisites compatible with image rendering. Finally, we present a theoretical toolbox allowing to determine the convergence speed of a sampling method from its spectral profile. This toolbox is used specifically to give indications about the design principles necessary for new sampling algorithms.