Added skeleton discussion

Signed-off-by: Jim Martens <github@2martens.de>

body.tex

@@ -735,4 +735,29 @@ detection is out of the question under theses circumstances.
 
 \chapter{Discussion}
 
+To recap, the hypothesis is repeated here.
+
+\begin{description}
+    \item[Hypothesis] Novelty detection using auto-encoders delivers similar or better object detection performance under open set conditions while being less computationally expensive compared to dropout sampling.
+\end{description}
+
+Based on the reported results, no clear answer can be given to the
+research question; rather new questions emerge: "Can auto-encoders
+work on realistic data sets like COCO with multiple different classes
+in one image?" In other words: "Is my experience due to
+implementation issues or a general theoretical problem of
+auto-encoders?"
+
+Despite best efforts, the results of Miller et al.~\cite{Miller2018}
+could not be replicated. This does not show anything though.
+To disprove Miller's work, any and all possible ways to replicate
+their work must fail. Contrarily, one successful replication
+proves the ability to replicate. On the surface, both Miller et al.
+and I used the same weights, the same network, and the same
+data sets. Only difference of note: they used a Caffe implementation
+of SSD, for this thesis the Tensorflow implementation with eager mode
+was used.
+
+
+
 \chapter{Closing}