BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Asia/Tokyo
X-LIC-LOCATION:Asia/Tokyo
BEGIN:STANDARD
TZOFFSETFROM:+0900
TZOFFSETTO:+0900
TZNAME:JST
DTSTART:18871231T000000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260817T171532Z
LOCATION:Hall B5 (2)\, B Block\, Level 5
DTSTART;TZID=Asia/Tokyo:20241205T095600
DTEND;TZID=Asia/Tokyo:20241205T101000
UID:siggraphasia_SIGGRAPH Asia 2024_sess125_papers_634@linklings.com
SUMMARY:Neural Kernel Regression for Consistent Monte Carlo Denoising
DESCRIPTION:Qi Wang (State Key Laboratory of CAD&CG, Zhejiang University);
  Pengju Qiao (Institute of Software Chinese Academy of Sciences; State Key
  Laboratory of CAD&CG, Zhejiang University); Yuchi Huo (State Key Laborato
 ry of CAD&CG, Zhejiang University; Zhejiang University); Shiji Zhai (Insti
 tute of Computing Technology, Chinese Academy of Sciences); Zixuan Xie (In
 stitute of Computing Technology, Chinese Academy of Sciences; Zhejiang Lab
 ); Rengan Xie (State Key Laboratory of CAD&CG, Zhejiang University); Wei H
 ua (Zhejiang Lab); Hujun Bao (State Key Laboratory of CAD&CG, Zhejiang Uni
 versity; Zhejiang University); and Tao Liu (Shanghai Maritime University, 
 College of Transport & Communications)\n\nUnbiased Monte Carlo path tracin
 g that is extensively used in realistic rendering produces undesirable noi
 se, especially with low samples per pixel (spp). Recently, several methods
  have coped with this problem by importing unbiased noisy images and auxil
 iary features to neural networks to either predict a fixed-sized kernel fo
 r convolution or directly predict the denoised result. However, since it i
 s impossible to produce arbitrarily high spp images as the training datase
 t, the network-based denoising fails to produce high-quality images under 
 high spp. More specifically, network-based denoising is not consistent and
  does not converge to the ground truth as the sampling rate increases. On 
 the other hand, the post-correction estimators yield a blending coefficien
 t for a pair of biased and unbiased images influenced by image errors or v
 ariances to ensure the consistency of the denoised image. As the sampling 
 rate increases, the blending coefficient of the unbiased image converges t
 o 1, that is, using the unbiased image as the denoised results. However, d
 ue to the difficulty of accurately predicting image errors or variances wi
 th low spp, these estimators usually produce artifacts. To address the abo
 ve problems, we take advantage of both kernel-predicting methods and post-
 correction denoisers. Specifically, we propose a novel kernel-based denois
 er based on distribution-free kernel regression consistency theory, which 
 does not explicitly combine the biased and unbiased results but constrains
  the kernel bandwidth to produce consistent results under high spp. Meanwh
 ile, our kernel regression method explores the bandwidth optimization in t
 he robust auxiliary feature space instead of the noisy image space, which 
 leads to consistent high-quality denoising at both low and high spp. Exper
 iments demonstrate that our method outperforms existing denoisers in accur
 acy and consistency.\n\nRegistration Category: Full Access, Full Access Su
 pporter\n\nLanguage Format: English Language\n\nSession Chair: Wenzel Jako
 b (École Polytechnique Féderale de Lausanne (EPFL), NVIDIA)\n\n
URL:https://asia.siggraph.org/2024/program/?id=papers_634&sess=sess125
END:VEVENT
END:VCALENDAR