Thesis of Basile Fraboni

Self-normalized estimators and joint path construction for efficient multi-view rendering

Defense date: 15/12/2021

Advisor: Véronique Eglin
Coadvisor: Jean-Claude Iehl


Physically based light transport simulations have gradually become the standard approach in the image production industry. The Path Tracing algorithm and its variants are used for their ability to simulate complex lighting phenomena. However, these simulations require the exploration of all the paths connecting a light source to the camera sensor. The construction of these paths is a complex sequential process and it is often necessary to construct and evaluate a very large number of paths to achieve acceptable noise levels in the images. This is even more problematic when adding production effects such as motion blur, depth of field and volumetric rendering. Another important aspect is that we often need to compute multiple images of the same scene, for example when rendering stereo pairs; lenticular images, light field images and holographic stereograms to visualize multiple points of view; animated camera trajectories for virtual tours or rendering motion sequences. In this thesis, our goal is to accelerate the rendering of multiple viewpoints in a single simulation by exploiting the coherence between cameras. This is challenging because existing methods for reusing paths between multiple views may introduce visible bias into the images, and are not suitable for all the production effects, materials, surfaces and volumes that a scene may include. We develop a new unidirectional algorithm to jointly render multiple images of the same scene at once. We introduce new methods for transforming and reusing paths from one camera to another in the presence of participating media and for generating sub-paths that best contribute to a subset of observers, as well as a new Monte Carlo estimator for combining the contributions of these paths with low variance. We demonstrate on several scenes including complex geometry, complex materials, participating media, and production effects that this method effectively reduces noise compared to frame-by-frame computations at equivalent computation time.

Mr Drettakis GeorgeDirecteur(trice) de recherche INRIA Sophia-AntipolisRapporteur(e)
Mr Holzschuch NicolasDirecteur(trice) de rechercheINRIA GrenobleRapporteur(e)
Mr Boubekeur TamyDirecteur(trice) de rechercheAdobeExaminateur​(trice)
Mr Paulin MathiasProfesseur(e)Université Paul Sabatier de ToulouseExaminateur​(trice)
Mme Eglin VéroniqueProfesseur(e) INSA LyonDirecteur(trice) de thèse
Mr Iehl Jean-ClaudeMaître de conférenceUniversité Claude Bernard Lyon 1Co-encadrant(e)