final project update

handouts/final_project.tex

@@ -63,6 +63,9 @@ \section{Implementation ideas}
 \paragraph{Faster null-scattering using space partitioning.}
 Implement Yue et al.'s kd-tree data structure~\cite{Yue:2010:UAS} or uniform grid~\cite{Yue:2011:TOS} for better upper bound of the extinction coefficients. As a bonus, implement progressive null tracking~\cite{Misso:2023:PNT}. 
 
+\paragraph{Emissive volumes.}
+Add emissive volumes to your Homework 2 code. You might want to read Pixar's volume sampling paper from Villemin and Hery~\cite{Villemin:2013:PIF}.
+
 \paragraph{Efficient transmittance estimation.}
 Implement Kettunen et al.'s unbiased ray marching estimator~\cite{Kettunen:2021:URT}.
 You can extend your volumetric path tracer in homework 2 for this.
@@ -171,7 +174,7 @@ \section{Research project ideas}
 Current rendering denoisers are usually trained separately from the rendering algorithms (in particular, they are usually trained using a standard path tracer). Can we train the denoiser with the sampling algorithm together to make them compensate for each other? You might want to read Bako's Offline Deep Importance Sampling~\cite{Bako:ODI:2019}.
 
 \paragraph{Learned multiple importance sampling.}
-While optimal importance sampling~\cite{Kondapaneni:2019:OMI} is shown to be optimal variance-wise, it has a few drawbacks: 1) it is computationally expensive, 2) it is only optimal under linear combination of existing contributions, and 3) it is only optimal for unbiased rendering -- it does not optimize for the Mean Square Error (bias squared plus variance). Therefore, it is tempting to simply learn to do multiple importance sampling. Take a few features as input (e.g., PDFs, roughness, curvature, ambient occlusion), train a parametric function (does not need to be a neural network) to combine your rendering samples. For this project, you don't need to implement in a renderer to start, you can just try it on some test functions.
+While optimal multiple importance sampling~\cite{Kondapaneni:2019:OMI} is shown to be optimal variance-wise, it has a few drawbacks: 1) it is computationally expensive, 2) it is only optimal under linear combination of existing contributions, and 3) it is only optimal for unbiased rendering -- it does not optimize for the Mean Square Error (bias squared plus variance). Therefore, it is tempting to simply learn to do multiple importance sampling. Take a few features as input (e.g., PDFs, roughness, curvature, ambient occlusion), train a parametric function (does not need to be a neural network) to combine your rendering samples. For this project, you don't need to implement in a renderer to start, you can just try it on some test functions.
 
 \paragraph{Neural mutation for Metropolis light transport.}
 Neural networks have been shown to be helpful for importance sampling in a Monte Carlo path tracer~\cite{Muller:2019:NIS}.

handouts/refs.bib

@@ -300,6 +300,16 @@ @inproceedings{Burley:2012:PBS
     year = {2012}
 }
 
+@article{Villemin:2013:PIF,
+    author = {Ryusuke Villemin and Christophe Hery},
+    title = {Practical Illumination from Flames},
+    year = {2013},
+    journal = {Journal of Computer Graphics Techniques},
+    volume  = {2},
+    number = {2},
+    pages = {142--155}
+}          
+
 @inproceedings{Aila:2013:QMB,
     title = {On quality metrics of bounding volume hierarchies},
     author = {Aila, Timo and Karras, Tero and Laine, Samuli},