Machine Learning in Visual Effects
April 20, 2018
3pm, MSC W301
Halfdan Ingvarsson, PhD
Mens Simia Software
Abstract: Visual effects in movies have come a long way, in recent years, in both quality and fidelity, and subsequent believability. Much of this progress has been due increases in computational power available and advances in light transport algorithms using importance sampling to guide the generation of the visual imagery. Advances have been made in development of MIS (multiple-importance sampling) algorithms to increase the visual contribution of each sample over naive methods. However, sampling errors and the sheer amount of computational cost that goes into reducing them is still a major barrier to progress and represent a significant cost for modern movie productions.
In recent years, research in computer graphics has increasingly turned to machine learning, to investigate how it can be leveraged to increase this visual contribution even further and drive down the amount of time required to get similar visual fidelity as with naive methods. In this talk we will be giving a brief overview of several of those techniques being researched, along with an introduction to importance sampling as it relates to light transport using in computer graphics.
Bio: Halfdan Ingvarsson is a veteran software developer, with over 20 years of specializing in developing software for visual effects production. He's worked on the development of many of the industry's leading animation and special effects softwares; such as Softimage, Autodesk's Maya, and most recently, on the Academy Award-winning Houdini, from Side Effects Software. Prior to his work in visual effect software, he worked in visual effects production, at The Mill, Pison. and others.
He currently does consulting work through Mens Simia Software. He's an active member of the Association of Computing Machinery and its SIGGRAPH chapter, having chaired multiple SIGGRAPH conference sessions and sat on the conference jury for multiple years.