A Study on Resolving Dynamic analysis
using deep learning Framework Logical
Methods
Yogesh Kumar1*, Dr. Gaurav Khandelwal2
1 Research Scholar, University of Technology, Jaipur, Rajasthan
Email: yogkum841@gmail.com
2 Professor and Supervisor, University of Technology, Jaipur, Rajasthan
Abstract - Deep learning technologies have gained a great deal of interest in recent months for
demonstrating material removal capabilities in a variety of separation, categorization, as well
as other machine vision activities. The majority of these deep networks are built employing
convolution or completely convolution structures. Our suggested Dynamic and Static gestures
System guided by a hand gesture high dimensional data methodology We use a snapshot of a
human's thoracic spine to determine the person's depth and the size of the space around his
or her hands. In this research, we provide a novel object-based deep-learning framework for
semantic segmentation of extremely high-resolution satellite data. In specifically, we leverage
object-based prior record by augmenting a fully convolutional neural network's training
strategy with an anisotropic diffusion data pre-processing phase and an additional loss term.
The goal of this restricted framework is to ensure that similar visual data is assigned to the
same semantic class. Here, we employ intermediate steps based on traditional image
processing techniques to aid in the resolution of the subsequent problem of classification,
detection, or segmentation. Research shows that adding pre- and post-processing steps to a
deep learning pipeline can boost model performance over using only the network. Recent
advances in UQ algorithms used for deep learning are analysed, and their potential for use in
relevance feedback is addressed. It also emphasizes basic research difficulties and directions
linked with UQ. Quantifiably, man-made classes with more precise geometry, such as
buildings, benefited the most from our strategy, particularly along object borders, indicating
the developed approach's enormous potential. The purpose of this paper is to offer an
overview of the approaches used inside deep learning frameworks to either appropriately
prepare the input (pre-processing) or enhance the outcomes of the network output (post-
processing), with an emphasis on digital pathology image analysis.
Keywords - Deep learning technologies, dimensional data methodology, pre-processing, deep-
learning framework, image processing methods, UQ methods, machine vision activities.
1. INTRODUCTION
The current manufacturing industry need beings’ smart frameworks that concentrate on human
abilities so instead of forcing individuals to change to whichever technologies is available. Show of
support interfaces, in this environment, tend to be using person prior knowledge and thus are
essential for understandable interaction of individuals with digital devices. Confusion concepts are
relevant in AI situations such as concrete learning algorithms and active learning (AL). Uncertainty
arises when the test and testing dataset are incompatible, where information ambiguity occurs from
class overlapping or variation in the data; nonetheless, assessing knowing error seems substantially
more difficult than predicting ate ambiguity. This is 2 types of unpredictability: experimental
electronica uncertainty and epistemology doubts. (See Fig. 1)
Figure 1: A diagram demonstrating the unique characteristics of aleatoric and scientific
uncertainties
Our work seeks to develop a comprehensive, perception gesture recognition strategy that can be
employed for human interface tasks in social or electronics industries. Specialised sensor complying
per health standards such as ISO/TS Europe. Europeans are usually needed in industrial applications
where human safety is paramount. Unfortunately, the situations that demand cognitive perceiving in
business or in cultural circumstances are different. Monocular cameras offer features over specialist
or number of co sensing, for example being lightweight, inexpensive, system independent, and simple
to combine. This is desirable for business robotics assistance in establishments such as lodging, and
offices. As an outcome, we offer a single method for recognizing dynamic and static gestures in RGB
pictures streams.
According to survey studying gesture communicating, majority gestures employed in manufacturing
operations are technically basic, and thus no ou pas demeanor is utilized in element manipulating.
Semantic segmentation has received significant research and development emphasis since it is
critical in very many essential tasks of computer vision such as scene understanding, information
processing, item tracking and identification and others. Different methodologies have been employed
to improve categorization outcomes and create powerful, generic models that are not dependent on
the training data. Algorithms for deep learning currently give trying to cut outcomes in a variety of
classification standard datasets utilizing generally two different architecture.
Nowadays, learning algorithms to medical image processing were quickly developed with the aim of
providing diagnostic information, integrating alternative viewpoints, or minimizing personal intra-. The
recent increase both in growth is affected and faster processors has allowed the implementation of
Convolutional neural network (CNNs) to the particular circumstance of computer vision, wherein
computational complexity are performed to the input image. CNNs is the most commonly used deep
network, and they are trained on either the comprehensive view or on texture features, understanding
the important features by optimising a certain gradient descent.
1.1. Research objectives and main contributions
We observe that only about 250 studies exploring the employment of UQ in AI are produced in
different disciplines throughout 2011 and 2021. (e.g., computer visions, image processing, medical
image analysis, signal processing and natural language processing). In this systematic review, we
tried to include the plurality of related articles. As either a consequence, this study can serve as a
definitive guide for such reader in traversing this constantly growing research field.
To study the Deep Teaching Model.
The scientific limitations of UQ methods are emphasized.
1.2. Preliminaries
Instead to exploring ambiguity in detail, we describe the structure of neural network feed forward
(NNs) using Probabilistic modelling in this article.
(c) Gaussian mixture model
Figure 2: Three possible uncertain models and their various network architecture are
illustrated informally
1.3. Algorithmic Methods
This does provide a few really mathematical equations for conducting application assessment based
on literature and reporting purposes and leveraging inputs like Source Lines of Code (Sea lines of
communication), total count of operations to do, as well as additional cost objects (CD) such as
dialect, methodological approach, risk assessments, competence, and so. Many algorithm systems,
including such Model based designs, the Pascal model, and FP pirouetted models, have already been
constructed. These models must be used as outputs, as things and keep about some factors such as
code lines (LOC), intricacy, and so on, which become impossible to acquire in the beginning stages of
the a Dd project.
Framework is one of the algorithm schemes established in Hoffman. Shortcut is employed to compute
scheduling, volume, work, and length. Shortcut is composed of:
i) Fundamental,
ii) Transitional, and
iii) In-depth Model.
Basic Framework evaluates the energy for a Hp spanning as tiny to large in a rapid and unstructured
method. Introductory and detailed Model based techniques minimize memory footprint in the
endeavour in individual in conjunction to optimum project outcome. Ou pas models include: This
model was created in the early 1990s to help in cost analysis. Researchers have concentrated on the
several methods that use soft computers technique, including artificial neural systems (NN), fuzzy
inference system models, and evolutionary automated process implementation. Based on the trained
dataset, NN is competent of generalisations.
2. SENTIMENT ANALYSIS
Sentiment analysis is a method in the field of natural language processing. Sentiment analysis, often
known as opinion mining, is the technique of recognizing human emotions and thoughts. 3. Diagram
based on the categorization level of a particular document or sentence, it determines if the supplied
material is positive, negative, or neutral. There are numerous methodologies of sentiment analysis.
This activity is completed by determining the emotion or opinion of the subjective element contained
on the inside of a text. The methodologies used to classify a piece of literature are focused on the
ideas stated in it, which may be favourable, unfavourable, or indifferent. The text getting analysed
could be a sentence, a paper, or anything really.
Figure 3 Sentiment Classification Techniques
The approaches for sentiment classification include categorized as learning algorithms and
etymological roots methods. Similarly, the method of machine learning is divided into supervised and
unsupervised techniques. Support Vector Machines (SVM), Naive Bayes (NN), Nave Bayes (NB), and
Maximum Entropy (ME) approaches are the most commonly used in reinforcement methods.
Support Vector Machines (SVM): It is a classification algorithm, meaning indicates it is not
stochastic. Because this is an unsupervised machine learning technique, a significant volume of
learning information is required. It classifying the training examples that used a high energy as a
boundary to separate the learning data into groups.
Neural Networks (NN): This is also defined as a classification algorithm, which itself is mainly
categorized as an unsupervised method of machine learning.
Naïve Bayes (NB): That's a well classification technique. It uses likelihood function to classify
keywords into the appropriate categories.
……………….1
Maximum Entropy (ME): It's also a classification technique, often known as an exponent or sign
back classification model. This predictor extracts a subset of features from the input, linear integrating
them, and uses this sum like an exponent.
The main contributions to providing this method were quantitative linguistics analyses of the gathered
corpora comprising either positive or negative phrases and the development of an emotion
classification employing these gathered corpora. The attitude classification method was employed to
perform an experimental assessment on authentic small non-posts. All of these activities will be
performed automatically, with no human interference.
2.1. Deep Learning Models Based Sentiment
Analysis especially when it relates to the use of layers of artificial neural networks for learning tasks.
Deep learning is a sophisticated machine learning approach in the research field. It can acquire many
degrees of representation and representations from input, permitting it to tackle controlled and
supervised academic difficulties. One of active research topic in Natural Language Recognition is
narrative text.
There are numerous techniques for performing sentiment classification tasks, both in supervised and
unsupervised methods. Svm Classification Machines (SVM), Maximum Entropy, Nave Bayes, and
other supervised methods for machine learning are samples. Sentiment lexicons, assessment team,
and syntax trends are example of unstructured machine learning methods.
Deep learning for sentiment analysis is becoming increasingly popular. The benefits of using deep
learning models include higher performance and more accurate results on learning tasks. Deep
neural network approaches will extract and classify features from documents and short text. Figure 3
depicts the use of several deep learning models on sentiment analysis, which is described in this
literature review.
Figure 4 Transfer Learning Models In use in Sentiment Analysis
Figure 5 MLP Algorithm Structure
This research paper demonstrates sentiment polarity using deep learning tools. For sentiment
classification, a multilayer perceptron approach is formulated, and decision tree-based feature ranking
is provided for feature extraction. The proposed method is tested on the IMDb dataset presented on
Figure 4.
A convolution layer is an eat neural net having one or many hidden layers, as well as input and output
layers. Each node in a learning algorithm is a neuron with a non-linear input signal in strata apart from
the input neurons. We employ supervised learning to train this type of network. Figure 5 illustrates a
multi - layer perceptron neural layer neural net.
2.2. Personalized recommendation Methods Focus on Sentiment Analysis
The sentiments evaluation method is used to the recommendation system to create an attitude
analysis-based recommendation system. The recommender uses the attitude approach. the study as a
decision-making system.
The Diagnostic and clinical information will be collected location specific metadata on nodes and uses
sentiment polarity to the data to determine if the remark is favourable or negative. This system
produces score evaluation based on prior user experiences and learned automata.
3. REVIEW OF LITERATURE
Tölgyessy, M 2017 A human client focuses to a particular area, robot recognizes the pointing signal,
processes its convergence with encompassing planar surface and moves to the objective. A profundity
camera mounted on the skeleton is utilized. The client doesn't have to wear any additional dress or
markers. The plan incorporates fundamental numerical ideas, for example, changes between
coordinate frameworks and vector deliberation of elements required for basic route, which other
momentum research works misses.
Yang et al in 2013 proposed a Bayesian-Surmising based suggestion framework for online
interpersonal organizations. This suggestion gives improved results than trust load proposal and
cooperative sifting suggestion. This recommender framework involves contingent likelihood for working
out rating closeness, where content appraisals are imparted to companions.
Anto et al in 2016 proposed an opinion examination-based recommender framework that gives item
evaluating utilizing feeling examination. This framework furnishes more exact outcomes by contrasting
and different arrangement procedures like Credulous Bayes and SVM on extricated tweets from
twitter.
Lydia et al in 2017 proposed a recommender framework that utilizes suppositions or feeling of clients
in regards to the administrations given by E-shopping site. For characterizing the opinions given by the
client they utilized stochastic learning calculation. This recommender framework will find out and
doesn't permit counterfeit surveys by utilizing the Macintosh address of the framework.
4. RELATED WORK
4.1. Convolutional Networks for Semantic Segmentation (Patch-Based Learning)
Fix based models can remove convoluted highlights by joining ghostly and spatial data simultaneously.
By and large, these were the primary structures adjusted from the remote detecting local area for the
semantic division of exceptionally high-goal datasets. In, the creators analysed three distinct
approaches to taking advantage of different convolutional structures (PatreoNet, AlexNet, CaffeNet,
GoogLeNet, VGG ConvNets, and OverFeat ConvNets):
1. Preparing the designs without any preparation utilizing just the dataset of interest;
2. Utilizing pre-prepared convolutional organizations and calibrating them as indicated by the
dataset of interest; and
3. Utilizing a pre-prepared convolutional network as a component extractor and supplanting the
last SoftMax layer with a SVM classifier.
Different fix based convolutional designs (DenseNet121, InceptionV3, VGG19, Xception, ResNet50,
and InceptionResNetV2) are additionally investigated in for the effective planning of wetlands.
Likewise, unique metropolitan climate classifications are identified in by taking advantage of ResNet
and VGG. Comparative methodologies have additionally been investigated for crop ID from high goal
symbolism. In addition, the proposed fix-based strategy in utilizes an AlexNet-based pre-prepared
engineering coordinated with Spatial Pyramid Pooling (SPP) and Side Oversight (SS) procedures.
4.2. Completely Convolutional Organizations for Semantic Division (Pixel-Based Learning)
Completely convolutional networks right now convey cutting edge brings about a few semantic division
benchmark challenges. The model result is a two-channel volume, one channel connected with the
parallel structure non-building veil, and the other connected with contacting boundaries of building
cases. Essentially, managed the programmed multi-class land division issue utilizing a Component
Pyramid Organization whose encoder depends on the ResNet50 organization.
4.3. Object-Based Learning for Semantic Segmentation
Rather than taking advantage of crude picture highlights removed by a profound learning organization,
a few late investigations attempt to utilize the item/super-pixel data during semantic division errands.
Objects/super-pixels likewise serve the need of lower computational intricacy, which is normally
expected since for profound convolutional networks a great many boundaries should be tuned
consuming huge GPU and computer processor centre hours. The far-off district addresses a
considerably bigger encompassing piece of the article lastly the worldwide locale portrays the whole
picture scene.
5. MATERIALS AND METHOD
5.1. The Developed Object-Based Learning Framework
In the recently referenced related assessments, the thing information is fundamentally planned under a
pre-taking care of or post processing way. Thusly, the key motivation here was to design a significant
learning framework that could gainfully facilitate thing depictions. To do accordingly, we included
object-based limits by solidifying luxuriously chipped away at picture depictions as priors to an
additional hardship term in the significant mind association. In Figure 5.1, a representative graphical
portrayal of the made methodology is presented.
Figure 5.1: Representation of the proposed object-based profound learning system
The info picture is taken care of to a F-CNN engineering with encoding (green) and translating (yellow)
layers. During the pixel-based advancement method, the semantic misfortune is determined by
contrasting the organization yield and the reference information, while the item-based misfortune
obliges the semantic names to be something similar with the predominant mark inside every super
pixel. The two misfortunes are then joined together to create the last division map.
All the more explicitly, let us characterize a picture Is addressed by a bunch of patches {Pi} for I = 1, 2.
. . Z with comparing thick ground truth explanation Ss including a bunch of names {li} for I = 1, 2, K.
Both the picture and the comparing ground truth incorporate s = 1, 2. . . D pixels.
In our trials, we managed an order issue with multiple classes and consequently we utilized the
multiclass cross entropy for the enhancement of the multitude of models.
………...….2
To this end, let us consider, without loss of over-simplification, a bunch of articles or fragmented
districts {Ci} for I = 1, 2, . . ., M. Expecting that every one of the pixels inside an article ought to have
similar name, we plan our misfortune as.
…………………3
5.2. Details of Implementation
In this part, we give a short diagram of the plain fix based and pixel-based methods that were used to
differentiate the results and the proposed framework and direct a broad evaluation. All preliminaries
were coordinated using the Yard significant learning structure.
5.2.1. Patch-Based Learning
Three ordinarily used plans were completed, explicitly ConvNet, AlexNet and VGG-16, which give one
single li for each Pi. In particular, ConvNet has a by and large direct designing. There are 4 blocks of
layers: 2 convolutional and 2 totally related. The second convolutional block follows a comparable
model and 2 totally related layers follow to make the last request thing.
5.2.2. Pixel-Based Learning
Different totally convolutional models were attempted in this work and we present them in this part.
One max-pooling movement, which applies 2 × 2 channels with a stage of 2, is similarly associated
with each block. Thusly, the principal data volume is down analysed in a portion of different times.
Then, the data picture is gone through the decoder where up testing strategies occur in a symmetric
manner.
5.3. Dataset and Training Procedure
Vaihingen is comprised of 33 exceptionally high-goal photographs with a typical size of 2494 2064,
three open channels (Nir, Red, and Green), and a ground test distance of 9 cm (i.e., the genuine
ground esteem that compares to the distance of two nearby pixel places).
In the example of Potsdam, 38 ortho-corrected photographs with a size of 6000 6000 and a ground
test division of 5 cm are open. There are four ghastly channels accessible here: red, green, blue, and
infrared. It ought to be noticed that the various classifications in this dataset are not similarly adjusted,
that demonstrates that a few classifications (for instance, Structures) are undeniably more common
than others (e.g., Vehicles). Table 1 shows the level of each class in regard toward the preparation
pictures.
Table 1 shows the proportion of every modern notion in the training datasets.
Groups
Vaihingen
Potsdam
Impervious_surface
0.255
0.264
Buildings
0.265
0.254
Low_vegetation
0.461
0.461
Trees
0.221
0.233
Cars
0.464
0.465
Clutter
0.263
0.232
Figure 1: The needed training time in minutes, as well as the best epoch for each design, can
be seen for Vaihingen (a) and Dresden (a) (b).
Table 2 lists all of the important info for both datasets, including computing costs and the
number of epochs for each design.
(a)
Mins/Epoch
Optimal Epoch
Total mins
ConvNet
1
26
27
Alex Net
1.2
22
23.2
VGG-11
11
45
56
SegNet
15
21
36
U-Net
23
46
69
PCN-Net
42
13
55
OB_Net
44
41
85
OB_Unet
44
12
56
0.255 0.265
0.461
0.221
0.464
0.263
0.264 0.254
0.461
0.233
0.465
0.232
The total of pictures in the training sets
Vaihingen Potsdam
(b)
Mins/Epoch
Optimal Epoch
Total mins
SegNet
145
464
609
U-Net
116
133
249
OB_Snet
125
461
586
OB_UBnet
221
222
443
5.4. Quantitative Evaluation Metrics
We utilized four particular evaluative measures to survey its nature of the outcomes: in general
exactness, accuracy, review, and F1 score. They are completely expressed as TP (Genuine Up-
sides), FP (Bogus Up-sides), and FN (Misleading Negatives) (Misleading Negatives). On the off
chance that we have a class l, the quantity of pixels effectively perceived as l is TP. FP signifies the
quantity of pixels that were mistakenly distinguished as l. Last, FN addresses pixels that have a place
with l yet have been relegated to that other class by the model.
6. EXPERIMENTAL RESULTS AND DISCUSSION
6.1. Quantitative Evaluation
The success of the built device number of training (OB Snet, OB Unet, and OB FCN) here on
Forewarning datasets was comparable with those of gem (ConvNet, AlexNet, and VGG-16) and
screen wholeheartedly (SegNet, U-Net, and FCN-16) networks. Figure 6.1 depicts the quantitative
research of the obtained Generally Accuracy (left).
The three instrument designs that accomplished more prominent accuracy in Articulate (OB Snet and
OB Unet) additionally are applied to the ensuing dam dataset (Figure 6.2, right). More specifically:
Structures for Skin Learning: The quantifiable discoveries for Convolutional, Resnet50, and VGG-16
for the Valve actuator dataset are introduced in Figure 6.2.
Figure 6.2 : A, B, the derived All Agreement (OA) rates for such Followers to share (left) &
Potsdam (right) sets after using the created entity substantial progress and the contemporary
country (neither patch-based or image) optimum network.
6.2. Discussion
The accompanying discoveries can be drawn from the previously mentioned quantitative and
subjective assessment results. As a general rule, completely convolutional structures are steadier and
more effective than fix based ones for semantic division errands in exceptionally high-goal pictures. All
the more explicitly, contingent upon the model utilized in every circumstance, in general exactness
levels expanded by 5-10%. The thought up object-based learning methodology improved by and large
exactness rates and F1 score, as well as the resultant precision rates per class.
7. CONCLUSIONS
The review considered and fabricated a component profound learning model in view of the
coordination of AML disentanglement and a misfortune capability that obliges the growing experience
with occasion priors. The introduced strategy is general and might be utilized with any completely
convolutional profound organizations. StaDNet, a bound together system for synchronous ID of static
hands and dynamic top half movements, is recommended in this review. An extraordinary learning-
based profundity assessor idea is likewise presented, which gauges the distance between an
individual and their hands utilizing simply center and top 2D skeleton information. This property implies
that monocular pictures are satisfactory, and the recommended system doesn't need profundity
detecting. Finally, the technique could compel pixels from a comparative thing to be classed on the
connected prevalent class while keeping ridiculous and spatial properties. Considering the quantitative
evaluation, more noticeable overall and per-class accuracy rates were reached when stood out from
the forefront. Emotionally, the philosophy made more moderate and less uproarious results while
holding the general construction, estimation, object edges, and restricts even more effectively. Future
11
15
23
55
85
56
23.2
56
1.2
27
23.2
56
1.2
11
15
44
23
55
42
44
44
23
55
85
27
23.2
56
ConvNet
Alex Net
VGG-11
SegNet
U-Net
FCN-16
OB_Snet
OB_Unet
OB_FCN
Patch –Based Pixel -Based Object -Based
(A)
55
85
23.2
56
1.2
1.2
11
44
23
55
23
55
27
23.2
56
SegNet
U-Net
OB_Snet
OB_Unet
OB_FCN
Pixel -Based Object -Based
(B)
possibilities remember the robotization of learning loads for the misfortune capability, as well as the
reconciliation of different rearrangements scales and picture portrayals to handle scale space
concerns all the more really. Our posture driven hard spatial consideration technique centers StaDNet
around chest area postures to mirror enormous scope appendage movements and close by pictures to
show fragile hand/finger developments. Thus, StaDNet beats the past procedures on the dataset. The
gave weight introduction method settle the irregularity in class conveyance in the dataset, permitting
boundaries to be streamlined for every one of the 249 classes' motions.
REFERENCES
1. Amini, A. Soleimany, S. Karaman, D. Rus, Spatial uncertainty sampling for end-to-end
control, 2018, arXiv:1805.04829.
2. G. Wang, W. Li, M. Aertsen, J. Deprest, S. Ourselin, T. Vercauteren, Aleatoric uncertainty
estimation with test-time augmentation for medical image seg- mentation with convolutional
neural networks, Neurocomputing 338 (2019) 34–45.
3. Gleeson, B.; MacLean, K.; Haddadi, A.; Croft, E.; Alcazar, J. Gestures for industry intuitive
human-robot communication from human observation. In Proceedings of the 2013 8th
ACM/IEEE International Conference on Human-Robot Interaction (HRI), Tokyo, Japan, 3–6
February 2013; pp. 349–356.
4. H. Liu, R. Ji, J. Li, B. Zhang, Y. Gao, Y. Wu, F. Huang, Universal adversarial perturbation via
prior driven uncertainty approximation, in: Proceedings of the IEEE International Conference
on Computer Vision, 2019, pp. 2941–2949.
5. H.P. Do, Y. Guo, A.J. Yoon, K.S. Nayak, Accuracy, uncertainty, and adaptability of automatic
myocardial ASL segmentation using deep CNN, Magn. Reson. Med. 83 (5) (2020) 1863–
1874.
6. J. Tompson, R. Goroshin, A. Jain, Y. LeCun, C. Bregler, Efficient object local- ization using
convolutional networks, in: Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, 2015, pp. 648–656.
7. Jung, P.G.; Lim, G.; Kim, S.; Kong, K. A wearable gesture recognition device for detecting
muscular activities based on air-pressure sensors. IEEE Trans. Ind. Inform. 2015, 11, 485–
494.
8. K. Brach, B. Sick, O. Dürr, Single shot MC dropout approximation, 2020, arXiv preprint
arXiv:2007.03293.
9. L. Yu, S. Wang, X. Li, C.-W. Fu, P.-A. Heng, Uncertainty-aware self-assembling model for
semi-supervised 3D left atrium segmentation, in: D. Shen, T. Liu, T.M. Peters, L.H. Staib, C.
Essert, S. Zhou, P.-T. Yap, A. Khan (Eds.), Medical Image Computing and Computer
Assisted Intervention, Springer International Publishing, Cham, 2019, pp. 605–613.
10. Neverova, N.; Wolf, C.; Taylor, G.W.; Nebout, F. Multi-scale Deep Learning for Gesture
Detection and Localization. In Proceedings of the European Conference on Computer Vision,
Zurich, Switzerland, 6–12 September 2014; pp. 474–490.
11. O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for biomedical image
segmentation, 2015, arXiv:1505.04597.
12. P. McClure, N. Kriegeskorte, Representing inferential uncertainty in deep neural networks
through sampling, in: International Conference on Learning Representations, ICLR 2017-
Conference Track Proceedings, 2016.
13. Papadomanolaki, M.; Vakalopoulou, M.; Karantzalos, K. Patch-based deep learning
architectures for sparse annotated very high resolution datasets. In Proceedings of the 2017
Joint Urban Remote Sensing Event (JURSE), Dubai, UAE, 6–8 March 2017.
14. Park, H.S.; Jung, D.J.; Kim, H.J. Vision-based Game Interface using Human Gesture. In
Proceedings of the Pacific-Rim Symposium on Image and Video Technology, Hsinchu,
Taiwan, 10–13 December 2006; pp. 662–671.
15. Shelhamer, E.; Long, J.; Darrell, T. Fully Convolutional Networks for Semantic Segmentation.
In Proceedings of the 2015 IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), Boston, MA, USA, 7–12 June 2015; pp. 3431–3440
16. T. Nair, D. Precup, D.L. Arnold, T. Arbel, Exploring uncertainty measures in deep networks for
multiple sclerosis lesion detection and segmentation, Med. Image Anal. 59 (2020) 101557.
17. Tölgyessy, M.; Dekan, M.; Ducho ˇn, F.; Rodina, J.; Hubinsky, P.; Chovanec, L. Foundations
of visual linear human–robot interaction ` via pointing gesture navigation. Int. J. Soc. Robot.
2017, 9, 509–523.
18. Tölgyessy, M.; Hubinsky, P.; Chovanec, L.; Ducho ˇn, F.; Babinec, A. Controlling a group of
robots to perform a common task by ` gestures only. Int. J. Imaging Robot. 2017, 17, 1–13. 4.
19. Volpi, M.; Tuia, D. Dense Semantic Labeling of Subdecimeter Resolution Images with
Convolutional Neural Networks. IEEE Trans. Geosci. Remote Sens. 2017, 55, 881–893.
20. Y. Gal, Z. Ghahramani, Dropout as a Bayesian approximation: Representing model
uncertainty in deep learning, in: International Conference on Machine Learning, 2016, pp.
1050–1059
