Towards real-time realistic computer graphics using Neural Radiance Fields

Speaker: Jorge Condor

Abstract:
The field of neural rendering (pairing neural networks with traditional computer graphics techniques such as ray-tracing, path-tracing or volume rendering) has seen an explosion of interest in the past few years. Particularly, Neural Radiance Fields (NeRFs for short) have emerged as an efficient way of distilling and encoding a sparse set of data (i.e. images of a certain object or scene) into compact high-quality continuous representations (in our example, a refined 3D model of the object or scene with accurate color and illumination). While this has tremendous implications in the fields of computer vision and robotics, it can also be useful in computer graphics, as an extremely efficient way of encoding complex light transport effects that would otherwise need to be fully physically simulated when creating our realistic images. In this talk I will be talking about NeRF, its implications and show an example application where we exploited and extended the technique to reduce in up to two orders of magnitude the amount of time required to render scenes with extremely complex light sources, such as the grand chandeliers we can find in opera houses and theatres.

Biography:
Jorge Condor is a PhD student in the Perception, Display and Fabrication Group at IDSIA-USI under the supervision of Prof. Piotr Didyk. His research lies in the intersection of computer graphics, machine learning and human perception, where he also combines them to solve problems in other fields such as additive manufacturing. Previously, he did two research internships in the Graphics and Imaging Lab (Zaragoza, Spain) under the supervision of Diego Gutiérrez and Adrián Jarabo, the last of which resulted in the paper he will be presenting during the talk.