Apparence multi-échelles pour le rendu réaliste et efficace des surfaces complexes

Efficient rendering of realistic and complex scenes is a major challenge for image synthesis. In the field of photorealistic rendering, the image is computed by simulating the physical interactions between the light and the content of the scene. Nowadays, production studios don't create simple scenes for their movies, but gigantic and extremely detailed worlds. This exponential rise of complexity is responsible for scalability problems. The amount of data and the complexity of the interactions are such that computing the image has become unreasonably costly, and this even on the most powerful computers. In the eld of real-time
rendering, the objective of photorealism is less pushed on, but the problem posed by the complexity resides in the amount of subpixel details. They are the source of artefacts like aliasing, popping and inconsistent changes of appearances. Subpixel details in fluence appearance at the level of the pixel and cannot just be removed. Indeed, the transition from pixel to subpixel during a zoom requires recovering the emerging visual effects with high accuracy. Hence, using a multi-scale strategy, in which subpixel details are represented like surface material, is a natural solution to this scalability problem. However, existing multi-scale models are mostly empirical and approximate. They are not designed to handle complex scenes and fail at recovering the correct appearance ; they don't satisfy the requirements for photorealistic quality demanded in production. We claim that it is possible to design efficient and high-quality multi-scale rendering algorithms. This can be done by modeling rigorously the problem of filtering the appearance at the level of the pixels. We explore this idea in the case of detailed surfaces, for which we use and revisit the theoretical background provided by microfacet theory.
In the first part of this manuscript, we design the physical microfacet-based model, which describes the interactions between light and a multi-scale surface. We start by revisiting the foundations of microfacet theory. Starting from the conservation of the visible projected area, we derive and unify existing results. Then, we extend the theory and make it compatible with multi-scale rendering paradigms. The understanding acquired from the model also allows us to derive new results. For instance, we show how invariance properties can be used to derive results in the case of anisotropic microsurfaces, for which anisotropy is due to stretched micro-geometry, the typical case which occurs with animated surfaces. We also propose a metric that we use to validate microfacet-based models.
In the second part, we show how to use these theoretical results to design multi-scale rendering applications for detailed surfaces. In practice, one needs to choose a representation for multi-scale surfaces and their materials, and compute how they interact with the incident light. We use a lightweight memory representation and show how to make it compatible with the animation of the models, so that there is no need to initialize it again for each image of an animated sequence. When the geometrical and optical attributes of the material are correlated, they have to be represented and evaluated in a way that depends on their visibility for the camera and the light source directions. We show that this problem can always be expressed as a color mapping problem. We derive its solution and an efficient way to evaluate it. Finally, we describe numerical integration schemes for the incident light dedicated to real-time and offline rendering. In particular, we propose a new importance sampling technique for microfacet-based materials that can be used in a typical Monte Carlo rendering.

Directors:

• Mr Fabrice Neyret (Directeur de Recherche - CNRS )

Raporteurs:

• Mr Jaroslav Krivanek (Professeur - Charles Unversity Prague )
• Mr Mathias Paulin (Professeur - Université Paul Sabatier Toulouse )

Examinators:

• Mr François Sillion (Directeur de recherche - INRIA )
• Mr Sébastien Lagarde (Senior graphic programmer - DICE Stockholm )
• Mr Christophe Bourlier (Directeur de recherche - CNRS )
• Mr Jos Stam (Research scientist - Alias Wavefront Toronto )