this is for holding javascript data
Volker Strobel added section_Experiment_Texture_recognition_task__.tex
almost 8 years ago
Commit id: 9d19bcd556bb82ed195b72051909592315c9edee
deletions | additions
diff --git a/section_Experiment_Texture_recognition_task__.tex b/section_Experiment_Texture_recognition_task__.tex
new file mode 100644
index 0000000..5c0ff62
--- /dev/null
+++ b/section_Experiment_Texture_recognition_task__.tex
...
\section{Experiment -- Texture recognition task}
\label{sec:experiment1}
The classification algorithm was validated on a recorded in-flight
video. For creating the data set, 500 images were recorded for each of
32 logos, placed in different environments, as shown in
Figure~\ref{exp1setup}. The images did not contain motion blur and
were recorded in different heights ranging from 20\,cm to 150\,cm.
For the training, three different approaches were compared: (i)
training on recorded images, (ii) training on synthetic data and (iii)
training on a combination of real-world data and synthetic data. The
combination (iii) should give an accuracy increase due to the larger
training dataset. For generating synthetic data, each marker was
randomly rotated and put into different environments using the data
augmentation tool presented in
Section~\ref{sec:syntheticdatageneration}. For both training and
testing, the same camera model was used.
Additionally, 1000 images of random scenes without markers were added
to the training set, to ensure that the classifier exhibits a high
true negative accuracy.