Add IPCV2 warping and classification

src/year1/image-processing-and-computer-vision/module2/sections/_classification.tex

@@ -0,0 +1,105 @@
+\chapter{Image classification}
+
+
+\section{Supervised datasets}
+
+\begin{description}
+    \item[Dataset] \marginnote{Dataset}
+        Given a set of labeled data, it can be split into:
+        \begin{descriptionlist}
+            \item[Train set] $D^\text{train} = \{ (\text{x}_\text{train}^{(i)}, y_\text{train}^{(i)}) \mid i = 1, \dots, N \}$.
+            \item[Test set] $D^\text{test} = \{ (\text{x}_\text{test}^{(i)}, y_\text{test}^{(i)}) \mid i = 1, \dots, M \}$.
+        \end{descriptionlist}
+
+        It is assumed that the two sets contain i.i.d. samples drawn from the same unknown distribution.
+\end{description}
+
+
+\subsection{Modified NIST (MNIST)}
+
+\begin{minipage}{0.45\linewidth}
+    \centering
+    \includegraphics[width=0.9\linewidth]{./img/mnist.png}
+\end{minipage}
+\begin{minipage}{0.5\linewidth}
+    \begin{descriptionlist}
+        \item[Content] Handwritten digits from 0 to 9.
+        \item[Number of classes] 10.
+        \item[Train set size] 50k.
+        \item[Test set size] 10k.
+        \item[Image format] $28 \times 28$ grayscale.
+    \end{descriptionlist}
+\end{minipage}
+
+
+\subsection{CIFAR10}
+
+\begin{minipage}{0.45\linewidth}
+    \centering
+    \includegraphics[width=0.9\linewidth]{./img/cifar10.png}
+\end{minipage}
+\begin{minipage}{0.5\linewidth}
+    \begin{descriptionlist}
+        \item[Content] Objects of various categories.
+        \item[Number of classes] 10.
+        \item[Train set size] 50k.
+        \item[Test set size] 10k.
+        \item[Image size] $32 \times 32$ RGB.
+    \end{descriptionlist}
+\end{minipage}
+
+
+\subsection{CIFAR100}
+
+\begin{minipage}{0.45\linewidth}
+    \centering
+    \includegraphics[width=0.7\linewidth]{./img/cifar100.png}
+\end{minipage}
+\begin{minipage}{0.5\linewidth}
+    \begin{descriptionlist}
+        \item[Content] Objects of various categories.
+        \item[Number of classes] 100 (20 super-classed with 5 sub-classes).
+        \item[Train set size] 50k.
+        \item[Test set size] 10k.
+        \item[Image size] $32 \times 32$ RGB.
+    \end{descriptionlist}
+\end{minipage}
+
+
+\subsection{ImageNet 21k}
+
+\begin{descriptionlist}
+    \item[Content] Objects of various categories.
+    \item[Number of classes] 21k synsets from WordNet organized hierarchically.
+    \item[Dataset size] 14 millions.
+    \item[Image size] Variable resolution RGB. Average size of $400 \times 350$.
+\end{descriptionlist}
+
+\begin{figure}[H]
+    \centering
+    \includegraphics[width=0.85\linewidth]{./img/imagenet21k.png}
+\end{figure}
+
+
+\subsection{ImageNet 1k}
+
+\begin{minipage}{0.45\linewidth}
+    \centering
+    \includegraphics[width=\linewidth]{./img/imagenet1k.png}
+\end{minipage}
+\begin{minipage}{0.5\linewidth}
+    \begin{descriptionlist}
+        \item[Content] Objects of various categories.
+        \item[Number of classes] 1000.
+        \item[Train set size] $1.3$ millions.
+        \item[Validation set size] 50k.
+        \item[Test set size] 100k.
+        \item[Image size] Variable resolution RGB. Often resized to $256 \times 256$.
+    \end{descriptionlist}
+\end{minipage}
+
+\begin{remark}
+    Performance is usually measured as top-5 accuracy as making a single prediction might be ambiguous due to the fact that the images can contain multiple objects.
+\end{remark}
+
+