clara/proseminar
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Update on Overleaf.

Browse source
This commit is contained in:
uxwmp 2023-06-18 12:30:35 +00:00 • committed by node
parent 48998b7380
commit 257108290e
4 changed files with 41 additions and 19 deletions
Download patch file Download diff file Expand all files Collapse all files

20
presentation/modules/motivation.tex
View file

@ -23,14 +23,14 @@ with Differentiable Monte Carlo Raytracing [\cite{ACM:inverse_rendering}]\\
\begin{frame}{Inverse rendering}
\begin{itemize}
\item Conventional rendering: Synthesize an Image from a 3D scene
\item Inverse rendering is solving the inverse problem: Synthesize a 3D scene from images
\item 3D modelling can be hard and time consuming
\item Inverse problem: Synthesize a 3D scene from images
%\item 3D modelling can be hard and time consuming
\item Approach:
\begin{itemize}
\item Approximate the 3D scene (often very coarse)
\item Approximate the 3D scene
\item Render the approximation differentiably
\item Calculate the error between the render and the images
\item Use ADAM or comparable gradient descent algorithm to minimize this error
\item Calculate the error
\item Use a gradient descent algorithm to minimize this error
\end{itemize}
\end{itemize}
\end{frame}
@ -75,11 +75,11 @@ with Differentiable Monte Carlo Raytracing [\cite{ACM:inverse_rendering}]\\
\end{frame}
\begin{frame}{Adversarial image generation}
\begin{itemize}
\item Problem: Labeling training data is tedious and expensive\\
\item Problem: Labeling training data is tedious\\
$\implies$ We want to automatically generate training data
\item One solution: Generative adversarial networks. Let two neural nets "compete"; a generator and a classifier. (e.g. AutoGAN [\cite{DBLP:AutoGAN}])\\
$\implies$ Impossible to make semantic changes to the image (e.g. lighting) since no knowlege of the 3D scene exists
\item Different solution: Generate image using differentiable raytracing, use gradient descent to optimize the result image to fall into a specific class\\
\item One solution: Generative adversarial networks. (e.g. AutoGAN [\cite{DBLP:AutoGAN}])\\
$\implies$ Impossible to make semantic changes to the image (e.g. lighting)
\item Different solution: Use differentiable raytracing\\
$\implies$ Scene parameters can be manipulated
\end{itemize}
\end{frame}
@ -97,7 +97,7 @@ with Differentiable Monte Carlo Raytracing [\cite{ACM:inverse_rendering}]\\
\end{minipage}
\centering
\caption{Left: Original images, features are correctly identified.\\
Right: adversarial examples, silver car is not recognized and pedestrians are identified where there are none. Only semantic features (color, position, rotation) have been changed.}
Right: adversarial examples, missing/wrong identifications after only semantic changes}
\label{fig:adv_img_example}
\end{figure}
\end{center}