2019-07-28 14:50:50 +02:00
|
|
|
\clearpage
|
|
|
|
\section*{Abstract}
|
|
|
|
|
2019-09-25 12:48:54 +02:00
|
|
|
Object detection in open set conditions is an important part of
|
|
|
|
applying object detection to real world data sets. The key question
|
2019-09-25 12:52:25 +02:00
|
|
|
is how to identify unknown or novel data. Dropout sampling with entropy
|
|
|
|
thresholding is one possible avenue to answer the question.
|
|
|
|
|
|
|
|
This thesis compares vanilla SSD and SSD with dropout sampling on the MS COCO
|
|
|
|
data set. The hyper parameters confidence threshold, entropy threshold,
|
|
|
|
usage of non-maximum suppression, usage of dropout layers during
|
|
|
|
testing, and the dropout keep ratio are varied to create different scenarios.
|
|
|
|
Both results from macro and micro averaging are presented.
|
2019-09-25 12:48:54 +02:00
|
|
|
|
|
|
|
Results show a significant improvement of the object detection performance
|
|
|
|
with a higher confidence threshold. The usage of non-maximum suppression
|
2019-09-29 10:19:04 +02:00
|
|
|
improves performance as well. Usage of dropout layers during testing does
|
2019-09-25 12:48:54 +02:00
|
|
|
not provide a better result, a lower keep ratio decreases the performance.
|