BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//132.216.98.100//NONSGML kigkonsult.se iCalcreator 2.20.4//
BEGIN:VEVENT
UID:20260508T150305EDT-1498xe5cGC@132.216.98.100
DTSTAMP:20260508T190305Z
DESCRIPTION:Abstract\n\nCapturing and synthesizing the appearance of real w
 orld phenomena is a long-standing goal of computer graphics. The gamut of 
 rendering applications is remarkably large\, ranging from real-time visual
 ization of immersive environments in virtual reality to physics-based ligh
 t transport simulations for animated feature films. Despite tremendous pro
 gress in improving the workflow of digital artists\, the main bottleneck t
 o their productivity remains content creation\, as a vast amount of manual
  labor is still required to author photorealistic 3D scenes with detailed 
 geometry\, emission profiles and materials. This thesis investigates stati
 stical and machine learning-based methods to circumvent (parts of) this te
 dious creation process. We present three practical rendering techniques\, 
 each targeting complementary aspects of the rendering cycle.\n\nFirst\, we
  extend the framework of delayed rejection Markov chain Monte Carlo to pri
 mary sample space Metropolis light transport (MLT) and introduce a two-sta
 ge proposal mechanism that automatically balances local exploration and co
 mputational efficiency. Our method\, called delayed rejection Metropolis l
 ight transport (DRMLT)\, exploits prioritization by proposing bolder or le
 ss costly transitions first before falling back on more timid or expensive
  kernels upon failure. We show how our sampler is general and robust by de
 ploying it on radiometrically complex scenes\, showcasing improved converg
 ence over prior MLT-based techniques.\n\nSecond\, we propose a learning-ba
 sed Monte Carlo method to efficiently importance sample illumination produ
 cts (e.g.\, the product of environment lighting and material) using normal
 izing flows. Our neural product sampler composes a flow head warp with an 
 emitter tail warp: the small conditional head is represented by a neural s
 pline flow\, while the large unconditional tail is discretized per environ
 ment map and its evaluation is instant. We show that imbuing our model wit
 h an near-exact emitter warp is an effective inductive bias for neural pro
 duct sampling and demonstrate significant variance reduction over previous
  methods on a range of rendering applications.\n\nFinally\, we present neu
 ral geometric level of detail (NGLoD)\, an efficient neural representation
  that\, for the first time\, enables real-time rendering of high-fidelity 
 neural signed distance fields (SDFs) while achieving high reconstruction q
 uality. Here\, we represent implicit surfaces using an octree-based featur
 e volume which adaptively fits shapes with multiple discrete levels of det
 ail (LoDs) and enables continuous LoD with SDF interpolation. We further d
 evelop an efficient GPU-based algorithm to interactively render our neural
  SDF representation via sparse octree ray traversal. We show how NGLoD can
  represent 3D shapes in a compressed format with higher visual fidelity th
 an traditional methods.\n
DTSTART:20241118T170000Z
DTEND:20241118T190000Z
LOCATION:Room 603\, McConnell Engineering Building\, CA\, QC\, Montreal\, H
 3A 0E9\, 3480 rue University
SUMMARY:PhD defence of Joey Litalien – Statistical and Learning-based Metho
 ds for High-performance Rendering
URL:https://www.mcgill.ca/ece/channels/event/phd-defence-joey-litalien-stat
 istical-and-learning-based-methods-high-performance-rendering-360741
END:VEVENT
END:VCALENDAR