BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Asia/Hong_Kong
X-LIC-LOCATION:Asia/Hong_Kong
BEGIN:STANDARD
TZOFFSETFROM:+0800
TZOFFSETTO:+0800
TZNAME:HKT
DTSTART:19911015T033000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20251218T030656Z
LOCATION:Meeting Room S426+S427\, Level 4
DTSTART;TZID=Asia/Hong_Kong:20251216T104000
DTEND;TZID=Asia/Hong_Kong:20251216T105000
UID:siggraphasia_SIGGRAPH Asia 2025_sess118_papers_1824@linklings.com
SUMMARY:Nonlinear Noise2Noise for Efficient Monte Carlo Denoiser Training
DESCRIPTION:Andrew Tinits and Stephen Mann (University of Waterloo)\n\nThe
  Noise2Noise method allows for training machine learning-based denoisers w
 ith pairs of input and target images where both the input and target can b
 e noisy. This removes the need for training with clean target images, whic
 h can be difficult to obtain. However, Noise2Noise training has a major li
 mitation: nonlinear functions applied to the noisy targets will skew the r
 esults. This bias occurs because the nonlinearity makes the expected value
  of the noisy targets different from the clean target image. Since nonline
 ar functions are common in image processing, avoiding them limits the type
 s of preprocessing that can be performed on the noisy targets. Our main in
 sight is that certain nonlinear functions can be applied to the noisy targ
 ets without adding significant bias to the results. We develop a theoretic
 al framework for analyzing the effects of these nonlinearities, and descri
 be a class of nonlinear functions with minimal bias.\n\nWe demonstrate our
  method on the denoising of high dynamic range (HDR) images produced by Mo
 nte Carlo rendering, where generating high-sample count reference images c
 an be prohibitively expensive. Noise2Noise training can have trouble with 
 HDR images, where the training process is overwhelmed by outliers and perf
 orms poorly. We consider a commonly used method of addressing these traini
 ng issues: applying a nonlinear tone mapping function to the model output 
 and target images to reduce their dynamic range. This method was previousl
 y thought to be incompatible with Noise2Noise training because of the nonl
 inearities involved.  We show that certain combinations of loss functions 
 and tone mapping functions can reduce the effect of outliers while introdu
 cing minimal bias. We apply our method to an existing machine learning-bas
 ed Monte Carlo denoiser, where the original implementation was trained wit
 h high-sample count reference images. Our results approach those of the or
 iginal implementation, but are produced using only noisy training data.\n\
 nRegistration Category: Full Access, Full Access Supporter\n\nSession Chai
 r: Oliver Deussen (University of Konstanz)\n\n
END:VEVENT
END:VCALENDAR