2010 Apr 05 at 10:30
DC 1304
Kartic Subr, Computer Graphics Group, INRIA-Grenoble
The need for realistic visuals pervades a broad spectrum of applications such as entertainment, engineering, medicine, defense and education. Realistic images are generated, from digital models of scenes, by simulating the interaction of light up to the tolerance of human perception. The disproportionate inflation in detail and complexity of models as well as the growing number of accountable optical phenomena together pose formidable new challenges in generating realistic pictures. By the use of Monte Carlo integration over various domains, the problem of simulating the flow and interaction of light with digital models of scenes is transformed into a sequence of sampling problems.
First, I will describe our novel strategies for efficiently integrating light energy over some of these domains. In particular, I will detail our recent analysis of finite-aperture cameras in the Fourier domain and an algorithm that exploits the analysis to produce significant gain (of at least an order of magnitude) in computation cost for generating high quality, realistic images containing blur due to defocus. Addressing the needs of certain applications, that impose a strict time-budget, I will also briefly present our recent work on simulating light transport in certain scattering media by avoiding explicit simulation of the optical processes. Next, I will present a simple framework for objective comparison of images generated using Monte Carlo estimators, by the use of the well-known statistical tests of hypotheses.
Finally, I will discuss my ongoing work and present my overall research theme spanning two categories: Core problems in realistic image synthesis and alternate model representation schemes for efficient generation of pictures. I will briefly present some recent work on multi-scale image manipulation, as an initial step towards work in model abstraction.
Bio ------ Kartic is a post-doctoral researcher in the computer graphics group at INRIA-Grenoble. He received his Master's (2005) and PhD (2008) degrees in computer science from the University of California, Irvine. He worked with James Arvo (UCI) and Fredo Durand (MIT) on theoretical and practical aspects of sampling problems for efficient Monte Carlo integration in computer graphics. During his PhD, he enjoyed a variety of short stints interning at venues including Rhythm and Hues Studios (Los Angeles), NVIDIA Inc. (Santa Clara) and Columbia University (NY). His research interests span most aspects of realistic picture generation and manipulation. He is particularly interested in the analysis and development of algorithms and numerical techniques for generating compellingly realistic pictures.