For the next STORM seminar, Mathias Paulin will give a presentation titled : Learning how to sample the path space

Abstract

Monte-Carlo methods for light transport simulation rely on efficient sampling of the path-space.
In this seminar, we will recall how the light transport simulation problem could be formulated into a Monte Carlo estimator of the esperance of a random variable defined in the path space. The density distribution of paths in this space could then be approximated by conditional probability, as in the path tracing algorithm, or by a global density function learned during rendering, hence resulting in guided path-tracing algorithms. We will discuss about two approaches of learning this density. These two approaches are based on the same neural network architecture and differ only on the space in which the density is defined that is either the primary sample space or the N-D cartesian space.
This seminar will be based on the two following papers :
Learning to importance sample in primary sample space, Quan Zheng, Matthias Zwicker – EG 2019
https://quan-zheng.github.io/publication/impsamplepss19/
Neural importance sampling, Thomas Müller et al., ACM TOG – 2019
https://tom94.net/data/publications/mueller18neural/mueller18neural-v4.pdf

This seminar will take place on Friday 13/11 at 12:30pm in room Salle des Thèses at IRIT.

Sanitary instructions : Masks are mandatory for attendees and speaker must be at least 1 meter (2 seems better) away from the audience.