Mots-clés : Global Illumination, Gaussian processes, Video games Durée du projet : Stage de Master II ) - 6 months. Possibilité de continuer en thèse. Nombre maximal d'étudiants : 1 Places disponibles : 1 Interrogation effectuée le : 27 janvier 2021, à 04 heures 01

Description

The goal of the internship is to experiment with the interpolation of global illumination using a Gaussian Process latent variable model. The latent variables will be chosen so as to optimize the quality of the interpolation.

Global illumination usually requires a costly calculation (as opposed to direct illumination). Based on the intuition that the two are linked by a low dimensional operator (despite the infinite dimension of the space direct and global illumination potentially belong to), we want to find low dimensional approximations of this relationship.

Since linear parameterizations (using e.g. PCA) have some limits (in particular because they cause visual artifacts) we aim for non linear parameterizations. This naturally leads to looking into Gaussian processes.

The student will essentially need to

get familiar with the mathematical principles of global illumination and Gaussian processes;

get familiar with the existing Maverick’s rendering platform;

implement the precomputation of images as well as the Gaussian process training in image space;

implement the real-time rendering loop in image space;

create a nice proof-of-concept demo in a non trivial geometric scene;

(optional) implement an object-space–videogame compatible–version where the light is stored into textures

(optional) discuss and experiments the advantages of different function bases to generate coefficients for direct illumination;

The internship requires to be fluent in C++ and very confortable with the acquisition of new mathematical concepts.

Do not hesitate to come talk with me. If you have your own ideas about global illumination and what can be done, I'm open as well. mail: cyril.soler@inria.fr.