Abstract

The synthesis of photo-realistic visuals of virtual environments entails simulation of the physics of light up to the tolerance of human perception. The inputs to this simulation are typically discrete approximations of inherently continuous signals such as 3D geometry and 6D material reflectance distributions. The output of the simulation is generally a light-field, or more commonly video or images (slices of the light-field). The analysis and exploitation of structure in these multidimensional signals is key for efficient discrete representation, as well as tractable simulation. In particular, two fundamental types of sampling problems are encountered during physically-based image synthesis: sampling for reconstruction and sampling for integration. The exploitation of local signal structure is crucial for both sampling problems.

I will begin my talk by providing a broad overview of photo-realistic depiction, along with my observations of the related trends and challenges. I will then place a few of my existing work within a broad view of the space of solutions. Next, I will introduce the notion of local light-fields and describe how their frequency domain analysis reveals invaluable information that may be utilised for efficient simulation of defocus due to finite-aperture cameras. I will then briefly mention current work (under review) on generalising this analysis and its extension to motion blur due to finite-exposure cameras. Finally, I will explain the shortcomings of such frequency-domain analysis and propose a new approach— local extremal theory for multidimensional signal analysis. I will speculate, with the help of examples, that knowledge of such local signal structure would be useful in applications beyond photo-realistic depiction— for example, light-field acquisition.