You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

1289 lines
68 KiB

% body thesis file that contains the actual content
\chapter{Introduction}
The introduction will explain the wider context first, before
providing technical details.
\subsection*{Motivation}
Famous examples like the automatic soap dispenser, which does not
recognise the hand of a black person but dispenses soap when presented
with a paper towel, raise the question of bias in computer
systems~\cite{Friedman1996}. Related to this ethical question regarding
the design of so called algorithms is the question of
algorithmic accountability~\cite{Diakopoulos2014}.
Supervised neural networks learn from input-output relations and
figure out by themselves what connections are necessary for that.
This feature is also their Achilles heel: it makes them effectively
black boxes and prevents any answers to questions of causality.
However, these questions of causality are of enormous consequence when
results of neural networks are used to make life changing decisions:
Is a correlation enough to bring forth negative consequences
for a particular person? And if so, what is the possible defence
against math? Similar questions can be raised when looking at computer
vision networks that might be used together with so called smart
CCTV cameras to discover suspicious activity.
This leads to the need for neural networks to explain their results.
Such an explanation must come from the network or an attached piece
of technology to allow adoption in mass. Obviously this setting
poses the question, how such an endeavour can be achieved.
For neural networks there are fundamentally two types of tasks:
regression and classification. Regression deals with any case
where the goal for the network is to come close to an ideal
function that connects all data points. Classification, however,
describes tasks where the network is supposed to identify the
class of any given input. In this thesis, I will work with both.
\subsection*{Object Detection in Open Set Conditions}
\begin{figure}
\centering
\includegraphics[scale=1.0]{open-set}
\caption{Open set problem: the test set contains classes that
were not present during training time.
Icons in this image have been taken from the COCO data set
website (\url{https://cocodataset.org/\#explore}) and were
vectorised afterwards. Resembles figure 1 of Miller et al.~\cite{Miller2018}.}
\label{fig:open-set}
\end{figure}
More specifically, I will look at object detection in the open set
conditions (see figure \ref{fig:open-set}).
In non-technical words this effectively describes
the kind of situation you encounter with CCTV cameras or robots
outside of a laboratory. Both use cameras that record
images. Subsequently, a neural network analyses the image
and returns a list of detected and classified objects that it
found in the image. The problem here is that networks can only
classify what they know. If presented with an object type that
the network was not trained with, as happens frequently in real
environments, it will still classify the object and might even
have a high confidence in doing so. Such an example would be
a false positive. Any ordinary person who uses the results of
such a network would falsely assume that a high confidence always
means the classification is very likely correct. If they use
a proprietary system they might not even be able to find out
that the network was never trained on a particular type of object.
Therefore, it would be impossible for them to identify the output
of the network as false positive.
This reaffirms the need for automatic explanation. Such a system
should by itself recognise that the given object is unknown and
hence mark any classification result of the network as meaningless.
Technically there are two slightly different approaches that deal
with this type of task: model uncertainty and novelty detection.
Model uncertainty can be measured, for example, with dropout sampling.
Dropout layers are usually used only during training but
Miller et al.~\cite{Miller2018} use them also during testing
to achieve different results for the same image making use of
multiple forward passes. The output scores for the forward passes
of the same image are then averaged. If the averaged class
probabilities resemble a uniform distribution (every class has
the same probability) this symbolises maximum uncertainty. Conversely,
if there is one very high probability with every other being very
low this signifies a low uncertainty. An unknown object is more
likely to cause high uncertainty which allows for an identification
of false positive cases.
Novelty detection is another approach to solve the task.
In the realm of neural networks it is usually done with the help of
auto-encoders that solve a regression task of finding an
identity function that reconstructs the given input~\cite{Pimentel2014}. Auto-encoders have
internally at least two components: an encoder, and a decoder or
generator. The job of the encoder is to find an encoding that
compresses the input as good as possible while simultaneously
being as loss-free as possible. The decoder takes this latent
representation of the input and has to find a decompression
that reconstructs the input as accurate as possible. During
training these auto-encoders learn to reproduce a certain group
of object classes. The actual novelty detection takes place
during testing: given an image, and the output and loss of the
auto-encoder, a novelty score is calculated. For some novelty
detection approaches the reconstruction loss is exactly the novelty
score, others consider more factors. A low novelty
score signals a known object. The opposite is true for a high
novelty score.
\subsection*{Research Question}
Auto-encoders work well for data sets like MNIST~\cite{Deng2012}
but perform poorly on challenging real world data sets
like MS COCO~\cite{Lin2014}, complicating any potential comparison between
them and object detection networks like \gls{SSD}.
Therefore, a comparison between model uncertainty with a network like
SSD and novelty detection with auto-encoders is considered out of scope
for this thesis.
Miller et al.~\cite{Miller2018} used an \gls{SSD} pre-trained on COCO
without further fine-tuning on the SceneNet RGB-D data
set~\cite{McCormac2017} and reported good results regarding
open set error for an \gls{SSD} variant with dropout sampling and entropy
thresholding.
If their results are generalisable it should be possible to replicate
the relative difference between the variants on the COCO data set.
This leads to the following hypothesis: \emph{Dropout sampling
delivers better object detection performance under open set
conditions compared to object detection without it.}
For the purpose of this thesis, I will use the \gls{vanilla} \gls{SSD} (as in: the original SSD) as
baseline to compare against. In particular, \gls{vanilla} \gls{SSD} uses
a per-class confidence threshold of 0.01, an IOU threshold of 0.45
for the \gls{NMS}, and a top \(k\) value of 200. For this
thesis, the top \(k\) value was changed to 20 and the confidence threshold
of 0.2 was tried as well.
The effect of an entropy threshold is measured against this \gls{vanilla}
SSD by applying entropy thresholds from 0.1 to 2.4 inclusive (limits taken from
Miller et al.). Dropout sampling is compared to \gls{vanilla} SSD
with and without entropy thresholding.
\paragraph{Hypothesis} Dropout sampling
delivers better object detection performance under open set
conditions compared to object detection without it.
\subsection*{Reader's Guide}
First, chapter \ref{chap:background} presents related works and
provides the background for dropout sampling.
Afterwards, chapter \ref{chap:methods} explains how \gls{vanilla} \gls{SSD} works, how
Bayesian \gls{SSD} extends \gls{vanilla} SSD, and how the decoding pipelines are
structured.
Chapter \ref{chap:experiments-results} presents the data sets,
the experimental setup, and the results. This is followed by
chapter \ref{chap:discussion}, focusing on the discussion and closing.
Therefore, the contribution is found in chapters \ref{chap:methods},
\ref{chap:experiments-results}, and \ref{chap:discussion}.
\chapter{Background}
\label{chap:background}
This chapter will begin with an overview over previous works
in the field of this thesis. Afterwards the theoretical foundations
of dropout sampling will be explained.
\section{Related Works}
The task of novelty detection can be accomplished in a variety of ways.
Pimentel et al.~\cite{Pimentel2014} provide a review of novelty detection
methods published over the previous decade. They showcase probabilistic,
distance-based, reconstruction-based, domain-based, and information-theoretic
novelty detection. Based on their categorisation, this thesis falls under
reconstruction-based novelty detection as it deals only with neural network
approaches. Therefore, the other types of novelty detection will only be
briefly introduced.
\subsection{Overview over types of Novelty Detection}
Probabilistic approaches estimate the generative probability density function (pdf)
of the data. It is assumed that the training data is generated from an underlying
probability distribution \(D\). This distribution can be estimated with the
training data, the estimate is defined as \(\hat D\) and represents a model
of normality. A novelty threshold is applied to \(\hat D\) in a way that
allows a probabilistic interpretation. Pidhorskyi et al.~\cite{Pidhorskyi2018}
combine a probabilistic approach to novelty detection with auto-encoders.
Distance-based novelty detection uses either nearest neighbour-based approaches
(e.g. \(k\)-nearest neighbour \cite{Hautamaki2004})
or clustering-based approaches
(e.g. \(k\)-means clustering algorithm \cite{Jordan1994}).
Both methods are similar to estimating the
pdf of data, they use well-defined distance metrics to compute the distance
between two data points.
Domain-based novelty detection describes the boundary of the known data, rather
than the data itself. Unknown data is identified by its position relative to
the boundary. A common implementation for this are support vector machines
(e.g. implemented by Song et al. \cite{Song2002}).
Information-theoretic novelty detection computes the information content
of a data set, for example, with metrics like entropy. Such metrics assume
that novel data inside the data set significantly alters the information
content of an otherwise normal data set. First, the metrics are calculated over the
whole data set. Afterwards, a subset is identified that causes the biggest
difference in the metric when removed from the data set. This subset is considered
to consist of novel data. For example, Filippone and Sanguinetti \cite{Filippone2011} provide
a recent approach.
\subsection{Reconstruction-based Novelty Detection}
Reconstruction-based approaches use the reconstruction error in one form
or another to calculate the novelty score. This can be auto-encoders that
literally reconstruct the input but it also includes MLP networks which try
to reconstruct the ground truth. Pimentel et al.~\cite{Pimentel2014} differentiated
between neural network-based approaches and subspace methods. The first were
further differentiated between MLPs, Hopfield networks, autoassociative networks,
radial basis function, and self-organising networks.
The remainder of this section focuses on MLP-based works, a particular focus will
be on the task of object detection and Bayesian networks.
Novelty detection for object detection is intricately linked with
open set conditions: the test data can contain unknown classes.
Bishop~\cite{Bishop1994} investigated the correlation between
the degree of novel input data and the reliability of network
outputs, and introduced a quantitative way to measure novelty.
The Bayesian approach provides a theoretical foundation for
modelling uncertainty \cite{Ghahramani2015}.
MacKay~\cite{MacKay1992} provided a practical Bayesian
framework for backpropagation networks. Neal~\cite{Neal1996} built upon
the work of MacKay and explored Bayesian learning for neural networks.
However, these Bayesian neural networks do not scale well. Over the course
of time, two major Bayesian approximations were introduced: one based
on dropout and one based on batch normalisation.
Gal and Ghahramani~\cite{Gal2016} showed that dropout training is a
Bayesian approximation of a Gaussian process. Subsequently, Gal~\cite{Gal2017}
showed that dropout training actually corresponds to a general approximate
Bayesian model. This means every network trained with dropout is an
approximate Bayesian model. During inference the dropout remains active,
this form of inference is called Monte Carlo Dropout (MCDO).
Miller et al.~\cite{Miller2018} built upon the work of Gal and Ghahramani: they
use MC dropout under open-set conditions for object detection.
In a second paper \cite{Miller2018a}, Miller et al. continued their work and
compared merging strategies for sampling-based uncertainty techniques in
object detection.
Teye et al.~\cite{Teye2018} make the point that most modern networks have
adopted other regularisation techniques. Ioffe and Szeged~\cite{Ioffe2015}
introduced batch normalisation which has been adapted widely. Teye et al.
showed how batch normalisation training is similar to dropout and can be
viewed as an approximate Bayesian inference. Estimates of the model uncertainty
can be gained with a technique named Monte Carlo Batch Normalisation (MCBN).
Consequently, this technique can be applied to any network that utilises
standard batch normalisation.
Li et al.~\cite{Li2019} investigated the problem of poor performance
when combining dropout and batch normalisation: dropout shifts the variance
of a neural unit when switching from train to test, batch normalisation
does not change the variance. This inconsistency leads to a variance shift which
can have a larger or smaller impact based on the network used.
Non-Bayesian approaches have been developed as well. Usually, they compare with
MC dropout and show better performance.
Postels et al.~\cite{Postels2019} provided a sampling-free approach for
uncertainty estimation that does not affect training and approximates the
sampling at test time. They compared it to MC dropout and found less computational
overhead with better results.
Lakshminarayanan et al.~\cite{Lakshminarayanan2017}
implemented a predictive uncertainty estimation using deep ensembles.
Compared to MC dropout, it shows better results.
Geifman et al.~\cite{Geifman2018}
introduced an uncertainty estimation algorithm for non-Bayesian deep
neural classification that estimates the uncertainty of highly
confident points using earlier snapshots of the trained model and improves,
among others, the approach introduced by Lakshminarayanan et al.
Sensoy et al.~\cite{Sensoy2018} explicitely model prediction uncertainty:
a Dirichlet distribution is placed over the class probabilities. Consequently,
the predictions of a neural network are treated as subjective opinions.
In addition to the aforementioned Bayesian and non-Bayesian works,
there are some Bayesian works that do not quite fit with the rest but
are important as well. Mukhoti and Gal~\cite{Mukhoti2018}
contributed metrics to measure uncertainty for semantic
segmentation. Wu et al.~\cite{Wu2019} introduced two innovations
that turn variational Bayes into a robust tool for Bayesian
networks: first, a novel deterministic method to approximate
moments in neural networks which eliminates gradient variance, and
second, a hierarchical prior for parameters and an empirical Bayes
procedure to select prior variances.
\section{Background for Dropout Sampling}
\begin{table}
\centering
\caption{Notation for background}
\label{tab:notation}
\begin{tabular}{l|l}
symbol & meaning \\
\hline
\(\mathbf{W}\) & weights \\
\(\mathbf{T}\) & training data \\
\(\mathcal{N}(0, I)\) & Gaussian distribution \\
\(I\) & independent and identical distribution \\
\(p(\mathbf{W}|\mathbf{T})\) & probability of weights given
training data \\
\(\mathcal{I}\) & an image \\
\(\mathbf{q} = p(y|\mathcal{I}, \mathbf{T})\) & probability
of all classes given image and training data \\
\(H(\mathbf{q})\) & entropy over probability vector \\
\(\widetilde{\mathbf{W}}\) & weights sampled from
\(p(\mathbf{W}|\mathbf{T})\) \\
\(\mathbf{b}\) & bounding box coordinates \\
\(\mathbf{s}\) & softmax scores \\
\(\overline{\mathbf{s}}\) & averaged softmax scores \\
\(D\) & detections of one forward pass \\
\(\mathfrak{D}\) & set of all detections over multiple
forward passes \\
\(\mathcal{O}\) & observation \\
\(\overline{\mathbf{q}}\) & probability vector for
observation \\
%\(E[something]\) & expected value of something
%\(\overline{\mathbf{z}}, \mathbf{z}\) & latent space representation \\
%\(d_T, d_z\) & discriminators \\
%\(e, g\) & encoding and decoding/generating function \\
%\(J_g\) & Jacobi matrix for generating function \\
%\(\mathcal{T}\) & tangent space \\
%\(\mathbf{R}\) & training/test data changed to be on tangent space
\end{tabular}
\end{table}
This section will use the \textbf{notation} defined in table
\ref{tab:notation} on page \pageref{tab:notation}.
To understand dropout sampling, it is necessary to explain the
idea of Bayesian neural networks. They place a prior distribution
over the network weights, for example a Gaussian prior distribution:
\(\mathbf{W} \sim \mathcal{N}(0, I)\). In this example
\(\mathbf{W}\) are the weights and \(I\) symbolises that every
weight is drawn from an independent and identical distribution. The
training of the network determines a plausible set of weights by
evaluating the probability output (posterior) over the weights given
the training data \(\mathbf{T}\): \(p(\mathbf{W}|\mathbf{T})\).
However, this
evaluation cannot be performed in any reasonable
time. Therefore approximation techniques are
required. In those techniques the posterior is fitted with a
simple distribution \(q^{*}_{\theta}(\mathbf{W})\). The original
and intractable problem of averaging over all weights in the network
is replaced with an optimisation task, where the parameters of the
simple distribution are optimised over~\cite{Kendall2017}.
\subsubsection*{Dropout Variational Inference}
Kendall and Gal~\cite{Kendall2017} showed an approximation for
classfication and recognition tasks. Dropout variational inference
is a practical approximation technique by adding dropout layers
in front of every weight layer and using them also during test
time to sample from the approximate posterior. Effectively, this
results in the approximation of the class probability
\(p(y|\mathcal{I}, \mathbf{T})\) by performing multiple forward
passes through the network and averaging over the obtained Softmax
scores \(\mathbf{s}_i\), given an image \(\mathcal{I}\) and the
training data \(\mathbf{T}\):
\begin{equation} \label{eq:drop-sampling}
p(y|\mathcal{I}, \mathbf{T}) = \int p(y|\mathcal{I}, \mathbf{W}) \cdot p(\mathbf{W}|\mathbf{T})d\mathbf{W} \approx \frac{1}{n} \sum_{i=1}^{n}\mathbf{s}_i
\end{equation}
With this dropout sampling technique, \(n\) model weights
\(\widetilde{\mathbf{W}}_i\) are sampled from the posterior
\(p(\mathbf{W}|\mathbf{T})\). The class probability
\(p(y|\mathcal{I}, \mathbf{T})\) is a probability vector
\(\mathbf{q}\) over all class labels. Finally, the uncertainty
of the network with respect to the classification is given by
the entropy \(H(\mathbf{q}) = - \sum_i q_i \cdot \log q_i\).
\subsubsection*{Dropout Sampling for Object Detection}
Miller et al.~\cite{Miller2018} apply the dropout sampling to
object detection. In that case \(\mathbf{W}\) represents the
learned weights of a detection network like SSD~\cite{Liu2016}.
Every forward pass uses a different network
\(\widetilde{\mathbf{W}}\) which is approximately sampled from
\(p(\mathbf{W}|\mathbf{T})\). Each forward pass in object
detection results in a set of detections, each consisting of bounding
box coordinates \(\mathbf{b}\) and softmax score \(\mathbf{s}\).
The detections are denoted by Miller et al. as \(D_i =
\{\mathbf{s}_i,\mathbf{b}_i\}\). The detections of all passes are put
into a large set \(\mathfrak{D} = \{D_1, ..., D_2\}\).
All detections with mutual intersection-over-union scores (IoU)
of \(0.95\) or higher are defined as an observation \(\mathcal{O}_i\).
Subsequently, the corresponding vector of class probabilities
\(\overline{\mathbf{q}}_i\) for the observation is calculated by averaging all
score vectors \(\mathbf{s}_j\) in a particular observation
\(\mathcal{O}_i\): \(\overline{\mathbf{q}}_i \approx \overline{\mathbf{s}}_i = \frac{1}{n} \sum_{j=1}^{n} \mathbf{s}_j\). The label uncertainty
of the detector for a particular observation is measured by
the entropy \(H(\overline{\mathbf{q}}_i)\).
If \(\overline{\mathbf{q}}_i\), which I called averaged class probabilities,
resembles a uniform distribution the entropy will be high. A uniform
distribution means that no class is more likely than another, which
is a perfect example of maximum uncertainty. Conversely, if
one class has a very high probability the entropy will be low.
In open set conditions it can be expected that falsely generated
detections for unknown object classes have a higher label
uncertainty. A threshold on the entropy \(H(\overline{\mathbf{q}}_i)\) can then
be used to identify and reject these false positive cases.
% SSD: \cite{Liu2016}
% ImageNet: \cite{Deng2009}
% COCO: \cite{Lin2014}
% YCB: \cite{Xiang2017}
% SceneNet: \cite{McCormac2017}
\chapter{Methods}
\label{chap:methods}
This chapter explains the functionality of \gls{vanilla} SSD, Bayesian SSD, and the decoding pipelines.
\section{Vanilla SSD}
\begin{figure}
\centering
\includegraphics[scale=1.2]{vanilla-ssd}
\caption{The \gls{vanilla} \gls{SSD} network as defined by Liu et al.~\cite{Liu2016}. VGG-16 is the base network, extended with extra feature layers. These predict offsets to anchor boxes with different sizes and aspect ratios. Furthermore, they predict the
corresponding confidences.}
\label{fig:vanilla-ssd}
\end{figure}
Vanilla \gls{SSD} is based upon the VGG-16 network (see figure
\ref{fig:vanilla-ssd}) and adds extra feature layers. The entire
image (always size 300x300) is divided up into anchor boxes. During
training, each of these boxes is mapped to a ground truth box or
background. For every anchor box both the offset to
the object and the class confidences are calculated. The output of the
SSD network are the predictions with class confidences, offsets to the
anchor box, anchor box coordinates, and variance. The model loss is a
weighted sum of localisation and confidence loss. As the network
has a fixed number of anchor boxes, every forward pass creates the same
number of detections---8732 in the case of \gls{SSD} 300x300.
Notably, the object proposals are made in a single run for an image -
single shot.
Other techniques like Faster R-CNN employ region proposals
and pooling. For more detailed information on SSD, please refer to
Liu et al.~\cite{Liu2016}.
\section{Bayesian SSD for Model Uncertainty}
Networks trained with dropout are a general approximate Bayesian model~\cite{Gal2017}. As such, they can be used for everything a true
Bayesian model could be used for. The idea is applied to \gls{SSD} in this
thesis: two dropout layers are added to \gls{vanilla} SSD, after the layers fc6 and fc7 respectively (see figure \ref{fig:bayesian-ssd}).
\begin{figure}
\centering
\includegraphics[scale=1.2]{bayesian-ssd}
\caption{The Bayesian \gls{SSD} network as defined by Miller et al.~\cite{Miller2018}. It adds dropout layers after the fc6
and fc7 layers.}
\label{fig:bayesian-ssd}
\end{figure}
Motivation for this is model uncertainty: an uncertain model will
predict different classes for the same object on the same image across
multiple forward passes. This uncertainty is measured with entropy:
every forward pass results in predictions, these are partitioned into
observations, and subsequently their entropy is calculated.
A higher entropy indicates a more uniform distribution of confidences
whereas a lower entropy indicates a larger confidence in one class
and very low confidences in other classes.
\subsection{Implementation Details}
For this thesis, an \gls{SSD} implementation based on Tensorflow~\cite{Abadi2015} and
Keras\footnote{\url{https://github.com/pierluigiferrari/ssd\_keras}}
was used. It was modified to support entropy thresholding,
partitioning of observations, and dropout
layers in the \gls{SSD} model. Entropy thresholding takes place before
the per-class confidence threshold is applied.
The Bayesian variant was not fine-tuned and operates with the same
weights that \gls{vanilla} \gls{SSD} uses as well.
\section{Decoding Pipelines}
The raw output of \gls{SSD} is not very useful: it contains thousands of
boxes per image. Among them are many boxes with very low confidences
or background classifications, those need to be filtered out to
get any meaningful output of the network. The process of
filtering is called decoding and presented for the three variants
of \gls{SSD} used in the thesis.
\subsection{Vanilla SSD}
Liu et al.~\cite{Liu2016} used Caffe for their original SSD
implementation. The decoding process contains largely two
phases: decoding and filtering. Decoding transforms the relative
coordinates predicted by \gls{SSD} into absolute coordinates. At this point
the shape of the output per batch is \((batch\_size, \#nr\_boxes, \#nr\_classes + 12)\). The last twelve elements are split into
the four bounding box offsets, the four anchor box coordinates, and
the four variances; there are 8732 boxes.
\glslocalreset{NMS}
Filtering of these boxes is first done per class:
only the class id, confidence of that class, and the bounding box
coordinates are kept per box. The filtering consists of
confidence thresholding and a subsequent \gls{NMS}.
All boxes that pass \gls{NMS} are added to a
per image maxima list. One box could make the confidence threshold
for multiple classes and, hence, be present multiple times in the
maxima list for the image. Lastly, a total of \(k\) boxes with the
highest confidences is kept per image across all classes. The
original implementation uses a confidence threshold of \(0.01\), an
IOU threshold for \gls{NMS} of \(0.45\) and a top \(k\)
value of 200.
The \gls{vanilla} SSD
per-class confidence threshold and \gls{NMS} has one
weakness: even if \gls{SSD} correctly predicts all objects as the
background class with high confidence, the per-class confidence
threshold of 0.01 will consider predictions with very low
confidences; as background boxes are not present in the maxima
collection, many low confidence boxes can be. Furthermore, the
same detection can be present in the maxima collection for multiple
classes. In this case, the entropy threshold would let the detection
pass because the background class has high confidence. Subsequently,
a low per-class confidence threshold does not restrict the boxes
either. Therefore, the decoding output is worse than the actual
predictions of the network.
Bayesian \gls{SSD} cannot help in this situation because the network
is not actually uncertain.
SSD was developed with closed set conditions in mind. A well trained
network in such a situation does not have many high confidence
background detections. In an open set environment, background
detections are the correct behaviour for unknown classes.
In order to get useful detections out of the decoding, a higher
confidence threshold is required.
\subsection{Vanilla SSD with Entropy Thresholding}
Vanilla \gls{SSD} with entropy tresholding adds an additional component
to the filtering already done for \gls{vanilla} SSD. The entropy is
calculated from all \(\#nr\_classes\) softmax scores in a prediction.
Only predictions with a low enough entropy pass the entropy
threshold and move on to the aforementioned per class filtering.
This excludes very uniform predictions but cannot identify
false positive or false negative cases with high confidence values.
\subsection{Bayesian SSD with Entropy Thresholding}
Bayesian \gls{SSD} has the speciality of multiple forward passes. Based
on the information in the paper, the detections of all forward passes
are grouped per image but not by forward pass. This leads
to the following shape of the network output after all
forward passes: \((batch\_size, \#nr\_boxes \, \cdot \, \#nr\_forward\_passes, \#nr\_classes + 12)\). The size of the output
increases linearly with more forward passes.
These detections have to be decoded first. Afterwards,
all detections are thrown away which do not pass a confidence
threshold for the class with the highest prediction probability.
Additionally, all detections with a background prediction of 0.8 or higher are discarded.
The remaining detections are partitioned into observations to
further reduce the size of the output, and
to identify uncertainty. This is accomplished by calculating the
mutual IOU score of every detection with all other detections. Detections
with a mutual IOU score of 0.95 or higher are partitioned into an
observation. Next, the softmax scores and bounding box coordinates of
all detections in an observation are averaged.
There can be a different number of observations for every image which
destroys homogenity and prevents batch-wise calculation of the
results. The shape of the results is per image: \((\#nr\_observations,\#nr\_classes + 4)\).
Entropy is measured in the next step. All observations with too high
entropy are discarded. Entropy thresholding in combination with
dropout sampling should improve identification of false positives of
unknown classes. This is due to multiple forward passes and
the assumption that uncertainty in some objects will result
in different classifications in multiple forward passes. These
varying classifications are averaged into multiple lower confidence
values which should increase the entropy and, hence, flag an
observation for removal.
The remainder of the filtering follows the \gls{vanilla} \gls{SSD} procedure: per-class
confidence threshold, \gls{NMS}, and a top \(k\) selection
at the end.
\chapter{Experimental Setup and Results}
\label{chap:experiments-results}
This chapter explains the used data sets, how the experiments were
set up, and what the results are.
\section{Data Sets}
This thesis uses the MS COCO~\cite{Lin2014} data set. It contains
80 classes, from airplanes to toothbrushes many classes are present.
The images are taken by camera from the real world, ground truth
is provided for all images. The data set supports object detection,
keypoint detection, and panoptic segmentation (scene segmentation).
The data of any data set has to be prepared for use in a neural
network. Typical problems of data sets include, for example,
outliers and invalid bounding boxes. Before a data set can be used,
these problems need to be resolved.
For the MS COCO data set, all annotations were checked for
impossible values: bounding box height or width lower than zero,
\(x_{min}\) and \(y_{min}\) bounding box coordinates lower than zero,
\(x_{max}\) and \(y_{max}\) coordinates lower than or equal to zero, \(x_{min}\) greater than \(x_{max}\),
\(y_{min}\) greater than \(y_{max}\), image width lower than \(x_{max}\),
and image height lower than \(y_{max}\). In the last two cases the
bounding box width and height were set to (image width - \(x_{min}\)) and
(image height - \(y_{min}\)) respectively;
in the other cases the annotation was skipped.
If the bounding box width or height afterwards is
lower than or equal to zero the annotation was skipped.
SSD accepts 300x300 input images, the MS COCO data set images were
resized to this resolution; the aspect ratio was not kept in the
process. MS COCO contains landscape and portrait images with (640x480)
and (480x640) as the resolution. This led to a uniform distortion of the
portrait and landscape images respectively. Furthermore,
the colour channels were swapped from RGB to BGR in order to
comply with the \gls{SSD} implementation. The BGR requirement stems from
the usage of Open CV in SSD: the internal channel order for
Open CV is BGR.
For this thesis, weights pre-trained on the sub data set trainval35k of the
COCO data set were used. These weights were created with closed set
conditions in mind, therefore, they had to be sub-sampled to create
an open set condition. To this end, the weights for the last
20 classes were thrown away, making them effectively unknown.
All images of the minival2014 data set were used but only ground truth
belonging to the first 60 classes was loaded. The remaining 20
classes were considered "unknown" and no ground truth bounding
boxes for them were provided during the inference phase.
A total of 31,991 detections remains after this exclusion. Of these
detections, a staggering 10,988 or 34,3\% belong to the persons
class, followed by cars with 1,932 or 6\%, chairs with 1,791 or 5,6\%,
and bottles with 1,021 or 3,2\%. Together, these four classes make up
around 49,1\% of the ground truth detections. This shows a huge imbalance
between the classes in the data set.
\section{Experimental Setup}
This section explains the setup for the different conducted
experiments. Each comparison investigates one particular question.
As a baseline, \gls{vanilla} \gls{SSD} with the confidence threshold of 0.01
and a \gls{NMS} IOU threshold of 0.45 was used.
Due to the low number of objects per image in the COCO data set,
the top \(k\) value was set to 20. Vanilla \gls{SSD} with entropy
thresholding uses the same parameters; compared to \gls{vanilla} SSD
without entropy thresholding, it showcases the relevance of
entropy thresholding for \gls{vanilla} SSD.
Vanilla \gls{SSD} was also run with 0.2 confidence threshold and compared
to \gls{vanilla} \gls{SSD} with 0.01 confidence threshold; this comparison
investigates the effect of the per class confidence threshold
on the object detection performance.
Bayesian \gls{SSD} was run with 0.2 confidence threshold and compared
to \gls{vanilla} \gls{SSD} with 0.2 confidence threshold. Coupled with the
entropy threshold, this comparison reveals how uncertain the network
is. If it is very certain the dropout sampling should have no
significant impact on the result. Furthermore, in two cases the
dropout was turned off to isolate the impact of \gls{NMS}
on the result.
Both, \gls{vanilla} \gls{SSD} with entropy thresholding and Bayesian \gls{SSD} with
entropy thresholding, were tested for entropy thresholds ranging
from 0.1 to 2.4 inclusive as specified in Miller et al.~\cite{Miller2018}.
\section{Results}
Results in this section are presented both for micro and macro averaging.
In macro averaging, for example, the precision values of each class are added up
and then divided by the number of classes. Conversely, for micro averaging the
precision is calculated across all classes directly. Both methods have
a specific impact: macro averaging weighs every class the same while micro
averaging weighs every detection the same. They will be largely identical
when every class is balanced and has about the same number of detections.
However, in case of a class imbalance the macro averaging
favours classes with few detections whereas micro averaging benefits classes
with many detections.
This section only presents the results. Interpretation and discussion is found
in the next chapter.
\subsection{Micro Averaging}
\begin{table}[ht]
\begin{tabular}{rcccc}
\hline
Forward & max & abs OSE & Recall & Precision\\
Passes & \(F_1\) Score & \multicolumn{3}{c}{at max \(F_1\) point} \\
\hline
\gls{vanilla} \gls{SSD} - 0.01 conf & 0.255 & 3176 & 0.214 & 0.318 \\
\gls{vanilla} \gls{SSD} - 0.2 conf & \textbf{0.376} & 2939 & \textbf{0.382} & 0.372 \\
\gls{SSD} with Entropy test - 0.01 conf & 0.255 & 3168 & 0.214 & 0.318 \\
% entropy thresh: 2.4 for \gls{vanilla} \gls{SSD} is best
\hline
Bay. \gls{SSD} - no DO - 0.2 conf - no \gls{NMS} \; 10 & 0.209 & 2709 & 0.300 & 0.161 \\
no dropout - 0.2 conf - \gls{NMS} \; 10 & 0.371 & \textbf{2335} & 0.365 & \textbf{0.378} \\
0.9 keep ratio - 0.2 conf - \gls{NMS} \; 10 & 0.359 & 2584 & 0.363 & 0.357 \\
0.5 keep ratio - 0.2 conf - \gls{NMS} \; 10 & 0.325 & 2759 & 0.342 & 0.311 \\
% entropy thresh: 1.2 for Bayesian - 2 is best, 0.4 for 3
% 0.5 for Bayesian - 6, 1.4 for 7, 1.4 for 8, 1.3 for 9
\hline
\end{tabular}
\caption{Rounded results for micro averaging. \gls{SSD} with Entropy test and Bayesian \gls{SSD} are represented with
their best performing entropy threshold with respect to \(F_1\) score. Vanilla \gls{SSD} with Entropy test performed best with an
entropy threshold of 2.4, Bayesian \gls{SSD} without \gls{NMS} performed best for 1.0,
and Bayesian \gls{SSD} with \gls{NMS} performed best for 1.4 as entropy
threshold.
Bayesian \gls{SSD} with dropout enabled and 0.9 keep ratio performed
best for 1.4 as entropy threshold, the run with 0.5 keep ratio performed
best for 1.3 as threshold.}
\label{tab:results-micro}
\end{table}
\begin{figure}[ht]
\begin{minipage}[t]{0.48\textwidth}
\includegraphics[width=\textwidth]{ose-f1-all-micro}
\caption{Micro averaged \(F_1\) score versus open set error for each variant. Perfect performance is an \(F_1\) score of 1 and an absolute OSE of 0.}
\label{fig:ose-f1-micro}
\end{minipage}%
\hfill
\begin{minipage}[t]{0.48\textwidth}
\includegraphics[width=\textwidth]{precision-recall-all-micro}
\caption{Micro averaged precision-recall curves for each variant tested.}
\label{fig:precision-recall-micro}
\end{minipage}
\end{figure}
Vanilla \gls{SSD} with a per-class confidence threshold of 0.2 performs best (see
table \ref{tab:results-micro}) with respect to the maximum \(F_1\) score
(0.376) and recall at the maximum \(F_1\) point (0.382). In comparison, neither
the \gls{vanilla} \gls{SSD} variant with a confidence threshold of 0.01 nor the \gls{SSD} with
an entropy test can outperform the 0.2 variant. Among the \gls{vanilla} \gls{SSD} variants,
the 0.2 variant also has the lowest number of open set errors (2939) and the
highest precision (0.372).
The comparison of the \gls{vanilla} \gls{SSD} variants with a confidence threshold of 0.01
shows no significant impact of an entropy test. Only the open set errors
are lower but in an insignificant way. The rest of the performance metrics is
identical after rounding.
Bayesian \gls{SSD} with disabled dropout and without \gls{NMS}
has the worst performance of all tested variants (\gls{vanilla} and Bayesian)
with respect to \(F_1\) score (0.209) and precision (0.161). The precision is not only the worst but also significantly lower compared to all other variants.
In comparison to all variants with 0.2 confidence threshold, it has the worst recall (0.300) as well.
With 2335 open set errors, the Bayesian \gls{SSD} variant with disabled dropout and
enabled \gls{NMS} offers the best performance with respect
to open set errors. It also has the best precision (0.378) of all tested
variants. Furthermore, it provides the best performance among all variants
with multiple forward passes.
Dropout decreases the performance of the network, this can be seen
in the lower \(F_1\) scores, higher open set errors, and lower precision
values. Both dropout variants have worse recall (0.363 and 0.342) than
the variant with disabled dropout.
However, all variants with multiple forward passes have lower open set
errors than all \gls{vanilla} \gls{SSD} variants.
The relation of \(F_1\) score to absolute open set error can be observed
in figure \ref{fig:ose-f1-micro}. Precision-recall curves for all variants
can be seen in figure \ref{fig:precision-recall-micro}. Both \gls{vanilla} SSD
variants with 0.01 confidence threshold reach much higher open set errors
and a higher recall. This behaviour is expected as more and worse predictions
are included.
All plotted variants show a similar behaviour that is in line with previously
reported figures, such as the ones in Miller et al.~\cite{Miller2018}
\subsection{Macro Averaging}
\begin{table}[t]
\begin{tabular}{rcccc}
\hline
Forward & max & abs OSE & Recall & Precision\\
Passes & \(F_1\) Score & \multicolumn{3}{c}{at max \(F_1\) point} \\
\hline
\gls{vanilla} \gls{SSD} - 0.01 conf & 0.370 & 1426 & 0.328 & 0.424 \\
\gls{vanilla} \gls{SSD} - 0.2 conf & \textbf{0.375} & 1218 & \textbf{0.338} & 0.424 \\
\gls{SSD} with Entropy test - 0.01 conf & 0.370 & 1373 & 0.329 & \textbf{0.425} \\
% entropy thresh: 1.7 for \gls{vanilla} \gls{SSD} is best
\hline
Bay. \gls{SSD} - no DO - 0.2 conf - no \gls{NMS} \; 10 & 0.226 & \textbf{809} & 0.229 & 0.224 \\
no dropout - 0.2 conf - \gls{NMS} \; 10 & 0.363 & 1057 & 0.321 & 0.420 \\
0.9 keep ratio - 0.2 conf - \gls{NMS} \; 10 & 0.355 & 1137 & 0.320 & 0.399 \\
0.5 keep ratio - 0.2 conf - \gls{NMS} \; 10 & 0.322 & 1264 & 0.307 & 0.340 \\
% entropy thresh: 1.2 for Bayesian - 2 is best, 0.4 for 3
% entropy thresh: 0.7 for Bayesian - 6 is best, 1.5 for 7
% 1.7 for 8, 2.0 for 9
\hline
\end{tabular}
\caption{Rounded results for macro averaging. \gls{SSD} with Entropy test and Bayesian \gls{SSD} are represented with
their best performing entropy threshold with respect to \(F_1\) score. Vanilla \gls{SSD} with Entropy test performed best with an
entropy threshold of 1.7, Bayesian \gls{SSD} without \gls{NMS} performed best for 1.5,
and Bayesian \gls{SSD} with \gls{NMS} performed best for 1.5 as entropy
threshold. Bayesian \gls{SSD} with dropout enabled and 0.9 keep ratio performed
best for 1.7 as entropy threshold, the run with 0.5 keep ratio performed
best for 2.0 as threshold.}
\label{tab:results-macro}
\end{table}
\begin{figure}[ht]
\begin{minipage}[t]{0.48\textwidth}
\includegraphics[width=\textwidth]{ose-f1-all-macro}
\caption{Macro averaged \(F_1\) score versus open set error for each variant. Perfect performance is an \(F_1\) score of 1 and an absolute OSE of 0.}
\label{fig:ose-f1-macro}
\end{minipage}%
\hfill
\begin{minipage}[t]{0.48\textwidth}
\includegraphics[width=\textwidth]{precision-recall-all-macro}
\caption{Macro averaged precision-recall curves for each variant tested.}
\label{fig:precision-recall-macro}
\end{minipage}
\end{figure}
Vanilla \gls{SSD} with a per-class confidence threshold of 0.2 performs best (see
table \ref{tab:results-macro}) with respect to the maximum \(F_1\) score
(0.375) and recall at the maximum \(F_1\) point (0.338). In comparison, the SSD
with an entropy test slightly outperforms the 0.2 variant with respect to
precision (0.425). Additionally, this is the best precision overall. Among
the \gls{vanilla} \gls{SSD} variants, the 0.2 variant also has the lowest
number of open set errors (1218).
The comparison of the \gls{vanilla} \gls{SSD} variants with a confidence threshold of 0.01
shows no significant impact of an entropy test. Only the open set errors
are lower but in an insignificant way. The rest of the performance metrics is
almost identical after rounding.
The results for Bayesian \gls{SSD} show a significant impact of \gls{NMS} or the lack thereof: maximum \(F_1\) score of 0.363 (with NMS) to 0.226
(without NMS). Dropout was disabled in both cases, making them effectively a
\gls{vanilla} \gls{SSD} run with multiple forward passes.
With 809 open set errors, the Bayesian \gls{SSD} variant with disabled dropout and
without \gls{NMS} offers the best performance with respect
to open set errors. The variant without dropout and enabled \gls{NMS} has the best \(F_1\) score (0.363), the best
precision (0.420) and the best recall (0.321) of all Bayesian variants.
Dropout decreases the performance of the network, this can be seen
in the lower \(F_1\) scores, higher open set errors, and lower precision and
recall values. However, all variants with multiple forward passes have lower open set errors than all \gls{vanilla} SSD
variants.
The relation of \(F_1\) score to absolute open set error can be observed
in figure \ref{fig:ose-f1-macro}. Precision-recall curves for all variants
can be seen in figure \ref{fig:precision-recall-macro}. Both \gls{vanilla} SSD
variants with 0.01 confidence threshold reach much higher open set errors
and a higher recall. This behaviour is expected as more and worse predictions
are included.
All plotted variants show a similar behaviour that is in line with previously
reported figures, such as the ones in Miller et al.~\cite{Miller2018}
\subsection{Class-specific results}
As mentioned before, the data set is imbalanced with respect to its
classes: four classes make up roughly 50\% of all ground truth
detections. Therefore, it is interesting to see the performance
of the tested variants with respect to these classes: persons, cars,
chairs, and bottles. Additionally, the results of the giraffe class are
presented as these are exceptionally good, although the class makes up
only 0.7\% of the ground truth. With this share, it is below
the average of roughly 0.89\% for each of the 56 classes that make up the
second half of the ground truth.
In some cases, multiple variants have seemingly the same performance
but only one or some of them are marked bold. This is informed by
differences prior to rounding. If two or more variants are marked bold
they had the exact same performance before rounding.
\begin{table}[tbp]
\begin{tabular}{rccc}
\hline
Forward & max & Recall & Precision\\
Passes & \(F_1\) Score & \multicolumn{2}{c}{at max \(F_1\) point} \\
\hline
\gls{vanilla} \gls{SSD} - 0.01 conf & 0.460 & \textbf{0.405} & 0.532 \\
\gls{vanilla} \gls{SSD} - 0.2 conf & \textbf{0.460} & \textbf{0.405} & \textbf{0.533} \\
\gls{SSD} with Entropy test - 0.01 conf & 0.460 & 0.405 & 0.532 \\
% entropy thresh: 1.7 for \gls{vanilla} \gls{SSD} is best
\hline
Bay. \gls{SSD} - no DO - 0.2 conf - no \gls{NMS} \; 10 & 0.272 & 0.292 & 0.256 \\
no dropout - 0.2 conf - \gls{NMS} \; 10 & 0.451 & 0.403 & 0.514 \\
0.9 keep ratio - 0.2 conf - \gls{NMS} \; 10 & 0.447 & 0.401 & 0.505 \\
0.5 keep ratio - 0.2 conf - \gls{NMS} \; 10 & 0.410 & 0.368 & 0.465 \\
% entropy thresh: 1.2 for Bayesian - 2 is best, 0.4 for 3
% entropy thresh: 0.7 for Bayesian - 6 is best, 1.5 for 7
% 1.7 for 8, 2.0 for 9
\hline
\end{tabular}
\caption{Rounded results for persons class. \gls{SSD} with Entropy test and Bayesian \gls{SSD} are represented with
their best performing macro averaging entropy threshold with respect to \(F_1\) score.}
\label{tab:results-persons}
\end{table}
It is clearly visible that the overall trend continues in the individual
classes (see tables \ref{tab:results-persons}, \ref{tab:results-cars}, \ref{tab:results-chairs}, \ref{tab:results-bottles}, and \ref{tab:results-giraffes}). However, the two \gls{vanilla} \gls{SSD} variants with only 0.01 confidence
threshold perform better than in the averaged results presented earlier.
Only in the chairs class, a Bayesian \gls{SSD} variant performs better (in
precision) than any of the \gls{vanilla} \gls{SSD} variants. Moreover, there are
multiple classes where two or all of the \gls{vanilla} \gls{SSD} variants perform
equally well. When compared with the macro averaged results,
giraffes and persons perform better across the board. Cars have a higher
precision than average but lower recall values for all but the Bayesian
SSD variant without \gls{NMS} and dropout. Chairs and bottles perform
worse than average.
\begin{table}[tbp]
\begin{tabular}{rccc}
\hline
Forward & max & Recall & Precision\\
Passes & \(F_1\) Score & \multicolumn{2}{c}{at max \(F_1\) point} \\
\hline
\gls{vanilla} \gls{SSD} - 0.01 conf & 0.364 & \textbf{0.305} & 0.452 \\
\gls{vanilla} \gls{SSD} - 0.2 conf & 0.363 & 0.294 & \textbf{0.476} \\
\gls{SSD} with Entropy test - 0.01 conf & \textbf{0.364} & \textbf{0.305} & 0.453 \\
% entropy thresh: 1.7 for \gls{vanilla} \gls{SSD} is best
\hline
Bay. \gls{SSD} - no DO - 0.2 conf - no \gls{NMS} \; 10 & 0.236 & 0.244 & 0.229 \\
no dropout - 0.2 conf - \gls{NMS} \; 10 & 0.336 & 0.266 & 0.460 \\
0.9 keep ratio - 0.2 conf - \gls{NMS} \; 10 & 0.332 & 0.262 & 0.454 \\
0.5 keep ratio - 0.2 conf - \gls{NMS} \; 10 & 0.309 & 0.264 & 0.374 \\
% entropy thresh: 1.2 for Bayesian - 2 is best, 0.4 for 3
% entropy thresh: 0.7 for Bayesian - 6 is best, 1.5 for 7
% 1.7 for 8, 2.0 for 9
\hline
\end{tabular}
\caption{Rounded results for cars class. \gls{SSD} with Entropy test and Bayesian \gls{SSD} are represented with
their best performing macro averaging entropy threshold with respect to \(F_1\) score. }
\label{tab:results-cars}
\end{table}
\begin{table}[tbp]
\begin{tabular}{rccc}
\hline
Forward & max & Recall & Precision\\
Passes & \(F_1\) Score & \multicolumn{2}{c}{at max \(F_1\) point} \\
\hline
\gls{vanilla} \gls{SSD} - 0.01 conf & 0.287 & \textbf{0.251} & 0.335 \\
\gls{vanilla} \gls{SSD} - 0.2 conf & 0.283 & 0.242 & 0.341 \\
\gls{SSD} with Entropy test - 0.01 conf & \textbf{0.288} & \textbf{0.251} & 0.338 \\
% entropy thresh: 1.7 for \gls{vanilla} \gls{SSD} is best
\hline
Bay. \gls{SSD} - no DO - 0.2 conf - no \gls{NMS} \; 10 & 0.172 & 0.168 & 0.178 \\
no dropout - 0.2 conf - \gls{NMS} \; 10 & 0.280 & 0.229 & \textbf{0.360} \\
0.9 keep ratio - 0.2 conf - \gls{NMS} \; 10 & 0.274 & 0.228 & 0.343 \\
0.5 keep ratio - 0.2 conf - \gls{NMS} \; 10 & 0.240 & 0.220 & 0.265 \\
% entropy thresh: 1.2 for Bayesian - 2 is best, 0.4 for 3
% entropy thresh: 0.7 for Bayesian - 6 is best, 1.5 for 7
% 1.7 for 8, 2.0 for 9
\hline
\end{tabular}
\caption{Rounded results for chairs class. \gls{SSD} with Entropy test and Bayesian \gls{SSD} are represented with
their best performing macro averaging entropy threshold with respect to \(F_1\) score. }
\label{tab:results-chairs}
\end{table}
\begin{table}[tbp]
\begin{tabular}{rccc}
\hline
Forward & max & Recall & Precision\\
Passes & \(F_1\) Score & \multicolumn{2}{c}{at max \(F_1\) point} \\
\hline
\gls{vanilla} \gls{SSD} - 0.01 conf & 0.233 & \textbf{0.175} & 0.348 \\
\gls{vanilla} \gls{SSD} - 0.2 conf & 0.231 & 0.173 & \textbf{0.350} \\
\gls{SSD} with Entropy test - 0.01 conf & \textbf{0.233} & \textbf{0.175} & 0.350 \\
% entropy thresh: 1.7 for \gls{vanilla} \gls{SSD} is best
\hline
Bay. \gls{SSD} - no DO - 0.2 conf - no \gls{NMS} \; 10 & 0.160 & 0.140 & 0.188 \\
no dropout - 0.2 conf - \gls{NMS} \; 10 & 0.224 & 0.170 & 0.328 \\
0.9 keep ratio - 0.2 conf - \gls{NMS} \; 10 & 0.220 & 0.170 & 0.311 \\
0.5 keep ratio - 0.2 conf - \gls{NMS} \; 10 & 0.202 & 0.172 & 0.245 \\
% entropy thresh: 1.2 for Bayesian - 2 is best, 0.4 for 3
% entropy thresh: 0.7 for Bayesian - 6 is best, 1.5 for 7
% 1.7 for 8, 2.0 for 9
\hline
\end{tabular}
\caption{Rounded results for bottles class. \gls{SSD} with Entropy test and Bayesian \gls{SSD} are represented with
their best performing macro averaging entropy threshold with respect to \(F_1\) score. }
\label{tab:results-bottles}
\end{table}
\begin{table}[tbp]
\begin{tabular}{rccc}
\hline
Forward & max & Recall & Precision\\
Passes & \(F_1\) Score & \multicolumn{2}{c}{at max \(F_1\) point} \\
\hline
\gls{vanilla} \gls{SSD} - 0.01 conf & \textbf{0.650} & \textbf{0.647} & \textbf{0.655} \\
\gls{vanilla} \gls{SSD} - 0.2 conf & \textbf{0.650} & \textbf{0.647} & \textbf{0.655} \\
\gls{SSD} with Entropy test - 0.01 conf & \textbf{0.650} & \textbf{0.647} & \textbf{0.655} \\
% entropy thresh: 1.7 for \gls{vanilla} \gls{SSD} is best
\hline
Bay. \gls{SSD} - no DO - 0.2 conf - no \gls{NMS} \; 10 & 0.415 & 0.414 & 0.417 \\
no dropout - 0.2 conf - \gls{NMS} \; 10 & 0.647 & 0.642 & 0.654 \\
0.9 keep ratio - 0.2 conf - \gls{NMS} \; 10 & 0.637 & 0.634 & 0.642 \\
0.5 keep ratio - 0.2 conf - \gls{NMS} \; 10 & 0.586 & 0.578 & 0.596 \\
% entropy thresh: 1.2 for Bayesian - 2 is best, 0.4 for 3
% entropy thresh: 0.7 for Bayesian - 6 is best, 1.5 for 7
% 1.7 for 8, 2.0 for 9
\hline
\end{tabular}
\caption{Rounded results for giraffe class. \gls{SSD} with Entropy test and Bayesian \gls{SSD} are represented with
their best performing macro averaging entropy threshold with respect to \(F_1\) score. }
\label{tab:results-giraffes}
\end{table}
\subsection{Qualitative Analysis}
% TODO: expand
This subsection compares \gls{vanilla} SSD
with Bayesian \gls{SSD} with respect to specific images that illustrate
similarities and differences between both approaches. For this
comparison, a 0.2 confidence threshold is applied. Furthermore, Bayesian
SSD uses \gls{NMS} and dropout with 0.9 keep ratio.
\begin{figure}
\begin{minipage}[t]{0.48\textwidth}
\includegraphics[width=\textwidth]{COCO_val2014_000000336587_bboxes_vanilla}
\caption{Image with stop sign and truck at right edge. Ground truth in blue, predictions in red, and rounded to three digits. Predictions are from \gls{vanilla} SSD.}
\label{fig:stop-sign-truck-vanilla}
\end{minipage}%
\hfill
\begin{minipage}[t]{0.48\textwidth}
\includegraphics[width=\textwidth]{COCO_val2014_000000336587_bboxes_bayesian}
\caption{Image with stop sign and truck at right edge. Ground truth in blue, predictions in red, and rounded to three digits. Predictions are from Bayesian \gls{SSD} with 0.9 keep ratio.}
\label{fig:stop-sign-truck-bayesian}
\end{minipage}
\end{figure}
The ground truth only contains a stop sign and a truck. The differences between \gls{vanilla} \gls{SSD} and Bayesian \gls{SSD} are almost not visible
(see figures \ref{fig:stop-sign-truck-vanilla} and \ref{fig:stop-sign-truck-bayesian}): the truck is neither detected by \gls{vanilla} nor Bayesian SSD, instead both detected a pottet plant and a traffic light. The stop sign is detected by both variants.
This behaviour implies problems with detecting objects at the edge
that overwhelmingly lie outside the image frame. Furthermore, the predictions are usually identical.
\begin{figure}
\begin{minipage}[t]{0.48\textwidth}
\includegraphics[width=\textwidth]{COCO_val2014_000000403817_bboxes_vanilla}
\caption{Image with a cat and laptop/TV. Ground truth in blue, predictions in red, and rounded to three digits. Predictions are from \gls{vanilla} SSD.}
\label{fig:cat-laptop-vanilla}
\end{minipage}%
\hfill
\begin{minipage}[t]{0.48\textwidth}
\includegraphics[width=\textwidth]{COCO_val2014_000000403817_bboxes_bayesian}
\caption{Image with a cat and laptop/TV. Ground truth in blue, predictions in red, and rounded to three digits. Predictions are from Bayesian \gls{SSD} with 0.9 keep ratio.}
\label{fig:cat-laptop-bayesian}
\end{minipage}
\end{figure}
Another example (see figures \ref{fig:cat-laptop-vanilla} and \ref{fig:cat-laptop-bayesian}) is a cat with a laptop/TV in the background on the right
side. Both variants detect a cat but the \gls{vanilla} variant detects a dog as well. The laptop and TV are not detected but this is expected since
these classes were not trained.
\chapter{Discussion and Outlook}
\label{chap:discussion}
First the results will be discussed, then possible future research and open
questions will be addressed.
\section*{Discussion}
The results clearly do not support the hypothesis: \textit{Dropout sampling delivers better object detection performance under open set conditions compared to object detection without it}. With the exception of open set errors, there
is no area where dropout sampling performs better than \gls{vanilla} SSD. In the
remainder of the section the individual results will be interpreted.
\subsection*{Impact of Averaging}
Micro and macro averaging create largely similar results. Notably, micro
averaging has a significant performance increase towards the end
of the list of predictions. This is signaled by the near horizontal movement
of the plot in both the \(F_1\) versus absolute open set error graph (see figure \ref{fig:ose-f1-micro}) and
the precision-recall curve (see figure \ref{fig:precision-recall-micro}).
This behaviour is caused by a large imbalance of detections between
the classes. For \gls{vanilla} \gls{SSD} with 0.2 confidence threshold there are
a total of 36,863 detections after \gls{NMS} and top \(k\).
The persons class contributes 14,640 detections or around 40\% to that number. Another strong class is cars with 2,252 detections or around
6\%. In third place come chairs with 1352 detections or around 4\%. This means that three classes have together roughly as many detections
as the remaining 57 classes combined.
In macro averaging, the cumulative precision and recall values are
calculated per class and then averaged across all classes. Smaller
classes quickly reach high recall values as the total number of
ground truth is small as well. The last recall and precision value
of the smaller classes is repeated to achieve homogenity with the largest
class. As a consequence, early on the average recall is quite high. Later on, only the values of the largest class still change which has only
a small impact on the overall result.
Conversely, in micro averaging the cumulative true positives
are added up across classes and then divided by the total number of
ground truth. Here, the effect is the opposite: the total number of
ground truth is very large which means the combined true positives
of 58 classes have only a smaller impact on the average recall.
As a result, the open set error rises quicker than the \(F_1\) score
in micro averaging, creating the sharp rise of open set error at a lower
\(F_1\) score than in macro averaging. The open set error
reaches a high value early on and changes little afterwards. This allows
the \(F_1\) score to catch up and produces the almost horizontal line
in the graph. Eventually, the \(F_1\) score decreases again while the
open set error further rises a bit.
Furthermore, the plotted behaviour implies that Miller et al.~\cite{Miller2018}
use macro averaging in their paper as the unique behaviour of micro
averaging was not reported in their paper.
\subsection*{Impact of Entropy}
There is no visible impact of entropy thresholding on the object detection
performance for \gls{vanilla} SSD. This indicates that the network has almost no
uniform or close to uniform predictions, the vast majority of predictions
has a high confidence in one class---including the background.
However, the entropy plays a larger role for the Bayesian variants---as
expected: the best performing thresholds are 1.0, 1.3, and 1.4 for micro averaging,
and 1.5, 1.7, and 2.0 for macro averaging. In all of these cases the best
threshold is not the largest threshold tested.
This is caused by a simple phenomenon: at some point most or all true
positives are in and a higher entropy threshold only adds more false
positives. Such a behaviour is indicated by a stagnating recall for the
higher entropy levels. For the low entropy thresholds, the low recall
is dominating the \(F_1\) score, the sweet spot is somewhere in the
middle. For macro averaging, it holds that a higher optimal entropy
threshold indicates a worse performance.
\subsection*{Non-Maximum Suppression and Top \(k\)}
\begin{table}[htbp]
\centering
\begin{tabular}{rccc}
\hline
variant & before & after & after \\
& entropy/NMS & entropy/NMS & top \(k\) \\
\hline
Bay. SSD, no dropout, no \gls{NMS} & 155,251 & 122,868 & 72,207 \\
no dropout, \gls{NMS} & 155,250 & 36,061 & 33,827 \\
\hline
\end{tabular}
\caption{Comparison of Bayesian \gls{SSD} variants without dropout with
respect to the number of detections before the entropy threshold,
after it and/or \gls{NMS}, and after top \(k\). The
entropy threshold 1.5 was used for both.}
\label{tab:effect-nms}
\end{table}
Miller et al.~\cite{Miller2018} supposedly did not use \gls{NMS}
in their implementation of dropout sampling. Therefore, a variant with disabled \glslocalreset{NMS}
\gls{NMS} was tested. The results are somewhat expected:
\gls{NMS} removes all non-maximum detections that overlap
with a maximum one. This reduces the number of multiple detections per
ground truth bounding box and therefore the false positives. Without it,
a lot more false positives remain and have a negative impact on precision.
In combination with top \(k\) selection, recall can be affected:
duplicate detections could stay and maxima boxes could be removed.
The number of observations was measured before and after the combination of entropy threshold and \gls{NMS} filter: both Bayesian \gls{SSD} without
NMS and dropout, and Bayesian \gls{SSD} with \gls{NMS} and disabled dropout
have the same number of observations everywhere before the entropy threshold. After the entropy threshold (the value 1.5 was used for both) and NMS, the variant with \gls{NMS} has roughly 23\% of its observations left
(see table \ref{tab:effect-nms} for absolute numbers).
Without \gls{NMS} 79\% of observations are left. Irrespective of the absolute
number, this discrepancy clearly shows the impact of \gls{NMS} and also explains a higher count of false positives:
more than 50\% of the original observations were removed with \gls{NMS} and
stayed without---all of these are very likely to be false positives.
A clear distinction between micro and macro averaging can be observed:
recall is hardly effected with micro averaging (0.300) but goes down equally with macro averaging (0.229). For micro averaging, it does
not matter which class the true positives belong to: every detection
counts the same way. This also means that top \(k\) will have only
a marginal effect: some true positives might be removed without \gls{NMS} but overall that does not have a big impact. With macro averaging, however,
the class of the true positives matters a lot: for example, if two
true positives are removed from a class with only few true positives
to begin with than their removal will have a drastic influence on
the class recall value and hence the overall result.
The impact of top \(k\) was measured by counting the number of observations
after top \(k\) has been applied: the variant with \gls{NMS} keeps about 94\%
of the observations left after NMS, without \gls{NMS} only about 59\% of observations
are kept. This shows a significant impact on the result by top \(k\)
in the case of disabled \gls{NMS}. Furthermore, some
classes are hit harder by top \(k\) then others: for example,
dogs keep around 82\% of the observations but persons only 57\%.
This indicates that detected dogs are mostly on images with few detections
overall and/or have a high enough prediction confidence to be
kept by top \(k\). However, persons are likely often on images
with many detections and/or have too low confidences.
In this example, the likelihood for true positives to be removed in
the person category is quite high. For dogs, the probability is far lower.
This is a good example for micro and macro averaging, and their impact on
recall.
\subsection*{Dropout Sampling and Observations}
\begin{table}[htbp]
\centering
\begin{tabular}{rccc}
\hline
variant & after & after \\
& prediction & observation grouping \\
\hline
Bay. SSD, no dropout, \gls{NMS} & 1,677,050 & 155,250 \\
keep rate 0.9, \gls{NMS} & 1,617,675 & 549,166 \\
\hline
\end{tabular}
\caption{Comparison of Bayesian \gls{SSD} variants without dropout and with
0.9 keep ratio of dropout with
respect to the number of detections directly after the network
predictions and after the observation grouping.}
\label{tab:effect-dropout}
\end{table}
The dropout variants have largely worse performance than the Bayesian variants
without dropout. This is expected as the network was not trained with
dropout and the weights are not prepared for it.
Gal~\cite{Gal2017}
showed that networks \textbf{trained} with dropout are approximate Bayesian
models. The Bayesian variants of \gls{SSD} implemented in this thesis are not fine-tuned or trained with dropout, therefore, they are not guaranteed to be such approximate models.
But dropout alone does not explain the difference in results. Both variants
with and without dropout have the exact same number of detections coming
out of the network (8732 per image per forward pass). With 16 images in a batch,
308 batches, and 10 forward passes, the total number of detections is
an astounding 430,312,960 detections. As such a large number could not be
handled in memory, only one batch is calculated at a time. That
still leaves 1,397,120 detections per batch. These have to be grouped into
observations, including a quadratic calculation of mutual IOU scores.
Therefore, these detections are filtered by removing all those with background
confidence levels of 0.8 or higher.
The number of detections per class was measured before and after the
detections were grouped into observations. To this end, the stored predictions
were unbatched and summed together. After the aforementioned filter
and before the grouping, roughly 0.4\% (in fact less than that) of the
more than 430 million detections are remaining (see table \ref{tab:effect-dropout} for absolute numbers). The variant with dropout
has slightly fewer predictions left compared to the one without dropout.
After the grouping, the variant without dropout has on average between
10 and 11 detections grouped into an observation. This is expected as every
forward pass creates the exact same result and these 10 identical detections
per \gls{vanilla} \gls{SSD} detection perfectly overlap. The fact that slightly more than
10 detections are grouped together could explain the marginally better precision
of the Bayesian variant without dropout compared to \gls{vanilla} SSD.
However, on average only three detections are grouped together into an
observation if dropout with 0.9 keep ratio is enabled. This does not
negatively impact recall as true positives do not disappear but offers
a higher chance of false positives. It can be observed in the results which
clearly show no negative impact for recall between the variants without
dropout and dropout with 0.9 keep ratio.
This behaviour implies that even a slight usage of dropout creates such
diverging anchor box offsets that the resulting detections from multiple
forward passes no longer have a mutual IOU score of 0.95 or higher.
\section*{Outlook}
The attempted replication of the work of Miller et al. raises a series of
questions that cannot be answered in this thesis. This thesis offers
one possible implementation of dropout sampling that technically works.
However, this thesis cannot answer why this implementation differs significantly
from Miller et al. The complete source code or otherwise exhaustive
implementation details of Miller et al. would be required to attempt an answer.
Future work could explore the performance of this implementation when used
on an \gls{SSD} variant that was fine-tuned or trained with dropout. In this case, it
should also look into the impact of training with both dropout and batch
normalisation.
Other avenues include the application to other data sets or object detection
networks.
To facilitate future work based on this thesis, the source code will be
made available and an installable Python package will be uploaded to the
PyPi package index. In the appendices can be found more details about the
source code implementation.