[4eyes] Talk today in the MAT Seminar Series

Mon Feb 5 10:27:26 PST 2018

Dear 4-Eyes Colleagues,

Sorry for the late notice, but I will be giving a high-level talk at the
MAT
seminar *today* at 1pm on our work on Monte Carlo denoising, if anyone
is interested.

See info below.

Best,

-Pradeep
---
Pradeep Sen
Associate Professor
UCSB MIRAGE Lab
Dept. of Electrical & Computer Engineering
University of California, Santa Barbara
Santa Barbara, CA 93106-9560

---------- Forwarded message ----------
From: MAT Seminars <mat-seminars at ucsb.edu>
Date: Sat, Feb 3, 2018 at 10:17 PM
Subject: [ Pradeep Sen - Feb. 5th - 1pm - THE MAT Seminar Series ]
To: mat-announce at lists.mat.ucsb.edu

<http://seminar.mat.ucsb.edu/>

<http://seminar.mat.ucsb.edu/>

<http://seminar.mat.ucsb.edu/>

MONTE CARLO (MC) PATH-TRACING
\
MONDAY. FEB.5TH. 2018 · Elings Hall (CNSI) · Room 1601 · 1PM

/
Monte Carlo (MC) path-tracing is now the most common rendering algorithm
used in industries ranging from architectural visualization to feature film
production. MC rendering systems produce photorealistic images by
simulating the physical flow of light through paths in the scene. However,
if too few light paths are computed, the images are filled with
objectionable noise, which made MC rendering unfeasible for over two
decades. Today, MC denoising algorithms are the most popular tool for
removing this noise, and they have been used in films such as Disney’s “Big
Hero 6” and featured in products such Pixar’s Renderman and NVIDIA's Iray.
However, it was only a few years ago that these post-process, screen-space
denoising approaches were considered unsuitable for tackling even small
amounts of MC noise, because it was thought they would either leave noise
artifacts or overblur scene detail.

In this talk, we will present a first-hand account of the MC denoising
revolution that has unfolded over the past decade and the key innovations
that made it possible. We begin in the summer of 2008, when we observed the
industrial “best-practices” for dealing with MC noise at Sony Pictures
Imageworks, one of the first studios to adopt a path-tracer as the primary
renderer. The limitations of the available approaches used in production
motivated us to start exploring the possibility of robust MC denoising
algorithms.

The first key idea we developed that we could output several features
computed at render time (sample positions, surface normals, texture values)
to make the denoiser more robust and effectively turn the rendering system
into a black box. Since these features often contained MC noise, we
realized we had to carefully adjust the manner in which these features were
used from pixel to pixel in order to remove the noise but preserve the
scene detail. The resulting system was the first to demonstrate that
high-quality, post-process general MC denoising was indeed possible.

In subsequent work, we observed that the problem could be modeled as a
supervised learning problem that would train a system to reproduce a
denoised output from noisy inputs. Since training a full denoiser given a
limited number of scenes was difficult, we trained an end-to-end system
that would output the parameters of a filter that would produce a result
comparable to the ground truth. Later, we extended this idea to work
compute the final color directly, allowing the denoiser to work robustly in
production environments. Our new system, developed in collaboration with
Disney and Pixar, was trained using millions of examples from the Pixar
film “Finding Dory” and then applied as a proof-of-concept to denoise the
renderings for the upcoming Pixar films Cars 3 and Coco, even though they
had completely different artistic styles and color palettes. Although MC
denoising has been credited as being one of two key “enabling technologies”
that brought path-tracing to feature film production, the journey is far
from over. We conclude the talk by discussing future directions for MC
denoising, and describe how it fits among the pantheon of tools available
for variance reduction.

/
Pradeep Sen is an Associate Professor in the UCSB MIRAGE Lab in the
Department of Electrical and Computer Engineering at the University of
California, Santa Barbara. He attended Purdue University from 1992 - 1996,
where he graduated with a B.S. in Computer and Electrical Engineering. He
then attended Stanford University where he received his M.S. in Electrical
Engineering in 1998 in the area of electron-beam lithography. In 2000, he
joined the Stanford Graphics Lab where he did research on real-time
rendering and computational photography. He received his Ph.D. in
Electrical Engineering in June 2006, advised by Dr. Pat Hanrahan. His
research interests include algorithms for image synthesis, computational
image processing, and computational photography, and he is a co-author of
over 30 technical publications, including ten SIGGRAPH/SIGGRAPH Asia/ToG
publications. Dr. Sen has been awarded more than $2.2 million in research
funding, including an NSF CAREER award in 2009.

seminar.mat.ucsb.edu
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cs.ucsb.edu/pipermail/ilab-users/attachments/20180205/859c2ad9/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: PradeepSen.gif
Type: image/gif
Size: 9729 bytes
Desc: not available
URL: <https://lists.cs.ucsb.edu/pipermail/ilab-users/attachments/20180205/859c2ad9/attachment-0001.gif>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: MAT_2018_logo_mail.png
Type: image/png
Size: 9174 bytes
Desc: not available
URL: <https://lists.cs.ucsb.edu/pipermail/ilab-users/attachments/20180205/859c2ad9/attachment-0001.png>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: MATSeminar_MailHeader.jpg
Type: image/jpeg
Size: 31625 bytes
Desc: not available
URL: <https://lists.cs.ucsb.edu/pipermail/ilab-users/attachments/20180205/859c2ad9/attachment-0002.jpg>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: portrait_PradeepSen.jpg
Type: image/jpeg
Size: 347393 bytes
Desc: not available
URL: <https://lists.cs.ucsb.edu/pipermail/ilab-users/attachments/20180205/859c2ad9/attachment-0003.jpg>