ELEKTROTEHNI ˇ SKI VESTNIK 91(1-2): 47–52, 2024
ORIGINAL SCIENTIFIC PAPER
Advancements in Gait Recognition: A Study on Gait
Energy Images and Gait Entropy Images
Stella Dumenˇ ci´ c
1
, Domagoj Pinˇ ci´ c
1
, Diego Suˇ sanj
1
, and
ˇ
Ziga Emerˇ siˇ c
2,† 1
University of Rijeka, Faculty of Engineering, Rijeka, Croatia
2
University of Ljubljana, Faculty of Computer and Information Science, Ljubljana, Slovenia
† E-mail: ziga.emersic@fri.uni-lj.si
Abstract. Gait recognition is a promising biometric modality due to its non-invasive nature and difficulty
to disguise. However, the performance still lacks compared to other, well established biometric modalities.
This paper presents results of our study on gait recognition, focusing on the comparison between Gait Energy
Images (GEI) and Gait Entropy Images (GEnI) under various conditions. Different methodologies are explored,
including deep learning techniques and Vision Transformers (ViTs), for feature extraction and classification.
The popular CASIA–B dataset is used to evaluate the performance across different walking conditions and
entropy measures. The effectiveness of gait recognition systems in accurately identifying individuals is shown,
thus highlighting the potential of GEnI in enhancing the recognition performance under varying conditions.
Keywords: Gait biometrics, GEI, Gait Entropy Image, CNN
Napredek pri prepoznavanju hoje: ˇ studija o energijskih
slikah hoje in entropijskih slikah hoje
Prepoznavanje hoje je obetavna biometriˇ cna modalnost zaradi
svoje neinvazivne narave in teˇ zav pri prirejanju. Zmogljivost je
ˇ se vedno slabˇ sa v primerjavi z drugimi, dobro uveljavljenimi
biometriˇ cnimi pristopi. Ta ˇ clanek predstavlja rezultate naˇ se
ˇ studije o prepoznavanju hoje, ki se osredotoˇ ca na primer-
javo med energijskimi slikami hoje (GEI) in entropijskimi
slikami hoje (GEnI) v razliˇ cnih pogojih. Predstavljena je
analiza razliˇ cnih metodologij, vkljuˇ cno s tehnikami globokega
uˇ cenja in transformatorji vida (ViT), za pridobivanje znaˇ cilk
in klasifikacijo. Uporabljena je priljubljena zbirka podatkov
CASIA–B za ocenjevanje zmogljivosti v razliˇ cnih pogojih hoje
in meritvah entropije. Prikazana je uˇ cinkovitost sistemov za
prepoznavanje hoje pri natanˇ cni identifikaciji posameznikov, s
ˇ cimer je poudarjen potencial GEnI pri izboljˇ sanju uˇ cinkovitosti
prepoznavanja v razliˇ cnih pogojih.
1 INTRODUCTION
Each individual has its own gait which describes its
unique way of walking. Unlike other biometric trait,
such as facial features, iris patterns, ear shapes and
fingerprints, the gait consists of several unique features.
One of these features is the greater distance to the sensor,
which does not require a direct interaction with a sensor
such as a camera. Moreover, the inherent difficulty of
altering the gait increases its reliability as a biometric
identifier and reduces the possibility of fraud. Extraction
of the gait data is possible even with low resolution
sensors, e.g. surveillance cameras. The use of gait recog-
nition includes the identification of individuals, which
Received 22 February 2024
Accepted 18 March 2024
can be used in access control mechanisms, surveillance
operations, and criminal investigations.
However, the implementation of gait biometrics in
real-life scenarios is hampered by several limitations and
challenges. First, the environmental factors that affect
the images, such as lighting changes, shadows and occlu-
sions, can significantly distort a person’s perceived gait,
similar to some other biometric modalities [1], [2], [3].
Second, different camera angles can result in different
appearances of the gait, even though the individual’s gait
signature. Common wearing modalities such as bags,
coats, hats or other accessories can visually alter a
person’s gait and thus complicate the interpretation of
gait recognition. In addition, the use of gait recognition
raises issues of privacy, bias and discrimination based
on physical characteristics or movement impairments.
Two methods for gait recognition are described in
the literature. The first method is based on compressing
silhouettes corresponding to a single gait cycle of an
individual into a consolidated image, resulting in a
representation of the gait features [4], [5]. Han et al. [4]
introduce the Gait Energy Image (GEI) to compresses
binary silhouettes extracted by background subtraction
from video images into a unified representation of the
gait. The second method considers the gait as a sequence
of individual silhouettes, each of which is used as
an input to a feature extractor [6], [7], [8], [9], [10],
[11]. Newer methods are mostly based on deep learning
techniques. Since their introduction in 2012 [12], Con-
volutional Neural Networks (CNNs) have significantly
influenced image-based deep learning and are now one
of the standards for gait recognition methodologies [6],

48 DUMEN
ˇ
CI
´
C ET. AL.
Classification
GEI
Traditional CNNs
Preprocessing
CASIA-B
GEnI
Vision Transformers
Preprocessing
Figure 1. Overview of the proposed evaluation pipeline of two different gait recognition approaches: GEI vs. GEnI.
[7], [13], [11].
Wu et al. [13] utilize CNNs for gait feature extraction
using similarity learning. GaitSet [6] propose a custom
CNN framework with triplet loss for learning gait rep-
resentations by treating a person’s silhouette as a set.
A detailed explanation of a standard gait recognition
pipeline is given in previous works [9], [11]. In recent
years, Vision Transformers (ViTs) have emerged as a
direct challenger to CNNs for image classification tasks.
Dosovitskiy et al. [14] introduce the ViT architecture.
The method employs a standard transformer encoder
used in natural language processing in the field of
computer vision, domain and specifically targets image
classification tasks. ViTs demonstrate a robust general-
ization capability [14]. In contrast to CNNs, ViTs require
fewer computational resources for training and at the
same time have better modeling capabilities.
Gait Entropy Images (GEnI) were developed as an
improvement of GEI. Their use helps to accumulating
the most significant motion information. GEnIs encap-
sulate a variance of pixel values in silhouette images
throughout an entire gait cycle into one image. In this
way, the motion data is preserved and remains resilient
to changes in covariate conditions that affect the appear-
ance. Bashir et al. [15] present GEnIs for capturing the
motion information and exhibiting resilience to changes
in the appearance. Extensive experiments show that
GEnIs outperform other methods, especially in scenarios
involving significant appearance changes. However, the
GEnI’s performance is susceptible to covariates that
directly affect the gait itself.
Rokanujjaman et al. [16] investigate gait signatures
by segmenting the human body into smaller fragments
and examining the effectiveness of combining these
fragments step by step. By using dynamically weighted
entropy-based gait representations as input features, their
approach outperforms both, the whole-based and the
part-based gait recognition methods. The results remain
consistent even with a subset of features, indicating the
robustness of the proposed approach and the importance
of specific parts. Jevan et al. [17] propose a novel
approach utilizing the Pal and Pal Entropy (GPPE)
features for gait recognition. The Principal Component
Analysis (PCA) is then used to extract salient features,
followed by training and testing with a Support Vector
Machine (SVM). Through rigorous experiments on both
the Treadmill dataset and the CASIA datasets A, B, and
C, the proposed method demonstrates a superior effec-
tiveness in gait representation, highlighting its potential
for a robust individual identification.
The contributions of this paper are:
• The utility of the Gait Entropy Images (GEnI)
over the traditional Gait Energy Images (GEI) in
handling appearance changes in gait recognition is
demonstrated.
• The performance of different entropy measures
(Shannon, Renyi, Tsallis) in the context of gait
recognition, with a focus on resilience to appear-
ance changes, is evaluated.
• The efficacy of two feature extraction methods, i.e.
the PCA-LDA and Vision Transformers (ViTs), in
enhancing the gait recognition accuracy, is com-
pared.
• Extensive experiments using the CASIA-B dataset
to assess the gait recognition performance across
various walking conditions and viewing angles are
conducted.
• Insights into the implications of using ViTs for
gait feature extraction, demonstrating their potential
over conventional methods in certain scenarios, are
provided.
2 MATERIALS AND METHODS
2.1 Datasets
The dataset used was CASIA-B [18]. It contains
124 subjects, three distinct walking conditions and 11
different camera viewing angles (from 0 to 180 with an
increment of 18°). The walking conditions are divided
into three categories: normal walking (NM), walking

ADV ANCEMENTS IN GAIT RECOGNITION: A STUDY ON GAIT ENERGY IMAGES AND GAIT ENTROPY IMAGES 49
while carrying a bag (BG) and walking with a coat or
jacket (CL). NM has six sequences per subject, while
BG and CL conditions have two sequences per subject.
Each subject has a total of 110 sequences, resulting in
a total number of nearly 1,118,000 silhouette images.
The dataset is divided into a train and a test subset
with the first 74 subjects used for training and the
remaining 50 subjects used for testing. The first four
sequences of the NM modality are assigned to the
gallery, while the remaining six NM sequences together
with the BG and CL sequences are assigned to the
queries. In our study, no differentiation is made between
the camera viewing angles.
2.2 Gait Energy Image
The standard image preprocessing methods [9], [19],
[20] are applied to the dataset used. First, the image
noise is filtered. Then the silhouettes for each subject
are extracted in a binary format, typically using methods
such as background subtraction. The images are then
standardized to ensure a uniform height and horizontal
alignment of all silhouettes. In the next step, the gait
cycle is estimated to generate a final representation of
the gait. The image-based gait features are used in the
form of GEIs [4]. GEIs capture the static features of a
gait sequence, for example the subject’s body shape, and
the dynamic aspects, including the frequency and phase
variations during locomotion. The GEI representationG
for a specific gait cycle is calculated using the following
formula:
G(i,j) =
1
N
N
X
t=1
I(i,j,t), (1)
where N is the number of the silhouette images in the
gait cycle, t is the image number at a specific point
in time within the gait cycle and I(i,j) is the original
silhouette image with the coordinates (i,j) in a 2D
image. Examples of the GEI representations for all three
conditions in CASIA-B are shown in Figure 2.
(a) NM (b) BG (c) CL
Figure 2. CASIA-B Gait Energy Images, with walking con-
ditions divided into three categories: normal walking (NM),
walking while carrying a bag (BG) and walking with a coat
or jacket (CL).
2.3 Gait Entropy Image
In the context of the size-normalized and centered sil-
houettes representing a gait cycle, the Shannon entropy
[21] is calculated for each pixel in the silhouette images.
It is used to quantify the uncertainty associated with
a random variable. By treating the intensity value of
the silhouettes at a specific pixel position as a discrete
random variable, the entropy of this variable over the
gait cycle is:
H
S
(i,j) = − K
X
k=1
p
k
(i,j)· log
2
(p
k
(i,j)), (2)
where i and j are the pixel coordinates and p
k
(i,j) is
the probability that the pixel takes on the K-th value
[15]. Since the silhouette images are binary, the number
of levels used for the entropy calculation is K = 2.
(a) Shannon, NM (b) Shannon, BG (c) Shannon, CL
(d) Renyi (α = 0.9),
NM
(e) Renyi (α = 0.9),
BG
(f) Renyi (α = 0.9),
CL
(g) Tsallis (q = 0.5),
NM
(h) Tsallis (q = 0.5),
BG
(i) Tsallis (q = 0.5),
CL
Figure 3. Resulting GEnI images for each entropy type and
walking condition. In the first column normal walking (NM),
in the middle column walking while carrying a bag (BG), and
in the last column walking with a coat or jacket (CL). In the
top row the Shannon entropy, in the middle row Renyi, and in
the bottom row Tsallis entropy types.
Since the Shannon-based GEnI images show an im-
provement compared to the GEI images, the effects of
other types of the entropy, namely the Renyi and Tsallis
entropy measures, are analyzed. All three entropies are
used for different problems and compared [22], [23],
[24], [25], [26], [27].
Both, the Renyi and Tsallis entropy measures are
generalizations of the Shannon entropy. The calculation
of the Renyi-based [28] GEnI image is performed as

50 DUMEN
ˇ
CI
´
C ET. AL.
follows:
H
R,α
(i,j) =
1
1− α · log
2
 
K
X
k=1
p
k
(i,j)
α !
, (3)
where α is the order of the entropy measure or a
parameter that determines its behavior.
The Tsallis entropy [29] is calculated using the equa-
tion:
H
T,q
(i,j) =
1
q− 1
· 
1− K
X
k=1
p
k
(i,j)
q
!
, (4)
whereq is the parameter that controls the degree of non-
extensivity and has a similar effect to the parameter α in the Renyi’s entropy.
For each entropy, the Gait Entropy Image G
E
(i,j)
is derived by fitting and discretizing H(i,j) to ensure
that its value falls in the range from 0 to 255, as given
below:
G
E
(i,j) =
(H(i,j)− H
min
)· 255
H
max
− H
min
, (5)
whereH
min
is the minimum calculated value andH
max
is the maximum calculated value for each GEnI image.
Since GEnI is calculated based on an entire gait cycle,
there are no issues with the temporal alignment. The
resulting GEnI images for each entropy type and walking
condition are shown in Figure 3.
2.4 Feature Extraction and Classification
After preprocessing the initial silhouette images and
converting them into GEI and GEnI images, the feature
extraction and classification are performed using two
feature extractors, i.e. the PCA-LDA combination and
the Visual Transformer based method.
After extracting GEI and GEnI from the video
sequence, the image data is converted into a one-
dimensional array by stacking the image columns on
top of each other. The dimensionality of this array is
reduced using the Principal Component Analysis (PCA)
followed by the Linear Discriminant Analysis (LDA).
This process is described in detail by Lenac et al.
[9] and Hofmann et al. [30]. Based on a preliminary
experimental assessment of the computation time, the
accuracy and our previous research, the number of
components to be extracted using PCA and LDA is set
to 50 and 10, respectively.
The second feature extractor is based on a self-
supervised learning paradigm for learning discrimina-
tive gait features. The DINO method is used, showing
promising results on various computer vision tasks such
as image classification and retrieval [31]. To adapt DINO
to the gait-specific data, the input sizes and augmenta-
tions used in training are modified. DINO demonstrates
the ability to segment objects in the foreground. This
is important in gait recognition scenarios where people
stand out from the background. Given the limited data
in gait datasets for training the ViT models from scratch,
a fine-tuning strategy is applied where DINO is first
trained on the ImageNet dataset and then fine-tuned for
the gait data. With this approach, DINO can be used
as a feature extractor to generate discriminative features
for subsequent classification tasks. A small ViT model
set up by Touvron et al. [32], with a patch size of 8, is
used.
PCA is applied to the training set. Both the gallery
and query are transformed into 50 features per image.
LDA is used to learn class-conditional densities from the
transformed gallery and further transforms the query into
10 features per image, which are used for the classifi-
cation. The DINO feature extractor is trained with the
images from the training subsets. Then the gait feature
for the gallery and query images is extracted and used
for the classification. The classification is performed
using the k-nearest neighbors (kNN) algorithm.
3 RESULTS AND DISCUSSION
The tests are performed with GEI and GEnI images. For
GEnI images, the Shannon, Renyi and Tsallis entropies
are used. For both entropies, different values ofα andq
are examined, ranging from 0.1 to 5.0. Table 1 compares
the results of the GEI and GEnI preprocessing ap-
proaches with two feature extraction extractor pipelines
for all three gait walking conditions.
GEI outperforms all GEnI approaches with the PCA-
LDA features. With ViT-8 it achieves the best result for
the BG condition. Tsallis GEnI with q = 1.5 achieves
the highest accuracy for the NM conditions with ViT-8.
Renyi with α = 0.9 proves to be the best for both the
CL conditions and overall accuracy with ViT-8.
Comparing the results between the entropies, shows
that the Shannon entropy does not perform best in any
of the categories. Tsallis performs best with PCA-LDA
and Renyi with ViT-8. In the overall ranking, GEnI based
on the Shannon entropy performs better than Renyi on
three different α values and better than Tsallis on two
different q values for the PCA-LDA feature extractor.
ViT-8 strongly favors Renyi and Tsallis compared to the
Shannon entropy and even GEI.
4 CONCLUSION
Our analysis reveals that Gait Energy Images (GEI)
excel when coupled with the traditional feature extrac-
tion methods such as PCA-LDA, affirming their com-
patibility with the established analytical frameworks.
Oppositely, Vision Transformers (ViTs) demonstrate a
notable preference for Gait Entropy Images (GEnI)
processed with the Renyi and Tsallis entropies. This
distinction suggests that ViTs’ advanced learning ca-
pabilities, particularly in recognizing and distinguishing
shapes, are better leveraged by the nuanced information

ADV ANCEMENTS IN GAIT RECOGNITION: A STUDY ON GAIT ENERGY IMAGES AND GAIT ENTROPY IMAGES 51
Image type α,q
PCA-LDA ViT-8
NM BG CL Overall NM BG CL Overall
GEI 90.45 66.48 40.82 65.92 99.18 82.24 25.27 68.90
GEnI: Shannon 88.27 55.83 35.91 60.01 99.18 77.86 31.09 69.38
GEnI: Renyi
0.5 85.73 57.93 35.82 59.82 98.73 79.04 31.91 69.89
0.7 87.91 57.66 36.09 60.55 98.73 79.22 32.09 70.01
0.9 87.91 55.75 36.00 59.88 98.91 78.68 32.82 70.14
1.5 87.64 53.37 33.18 58.07 99.18 78.14 31.09 69.47
GEnI: Tsallis
0.5 86.91 58.93 36.27 60.70 98.55 79.04 32.36 69.98
0.7 88.27 57.20 36.45 60.64 98.82 78.68 32.09 69.87
0.9 87.91 55.75 36.09 59.92 99.00 78.87 32.27 70.05
1.5 88.36 54.47 35.64 59.49 99.27 78.32 31.82 69.80
Table 1. Accuracy scores for both approaches for the feature extraction on GEI and GEnI input images for all walking conditions
of CASIA-B, where NM, BG and CL denote normal walking, walking while carrying a bag, and walking with a coat or jacket
respectively.
captured by the Renyi and Tsallis entropies. By provid-
ing a detailed representation of the variability around
the silhouette of a person, these entropies enhance the
model’s ability to discern subtle differences in the gait
patterns. Besides highlighting the evolving landscape of
the gait recognition technologies, this study also shows
a promising direction for a future research in optimizing
feature extraction techniques to leverage the strengths of
contemporary deep learning models.
REFERENCES
[1]
ˇ
Z. Emerˇ siˇ c, T. Ohki, M. Akasaka, T. Arakawa, S. Maeda,
M. Okano, Y . Sato, A. George, S. Marcel, I. Ganapathi et al.,
“The unconstrained ear recognition challenge 2023: Maximizing
performance and minimizing bias.”
[2] A. Hrovatiˇ c, P. Peer, V .
ˇ
Struc, and
ˇ
Z. Emerˇ siˇ c, “Efficient ear
alignment using a two-stack hourglass network,” IET Biometrics,
2023.
[3]
ˇ
Z. Emerˇ siˇ c, D. Suˇ sanj, B. Meden, P. Peer, and V .
ˇ
Struc, “Con-
texednet: Context–aware ear detection in unconstrained settings,”
IEEE Access, pp. 1–1, 2021.
[4] J. Han and B. Bhanu, “Individual recognition using gait energy
image,” IEEE transactions on pattern analysis and machine
intelligence, vol. 28, no. 2, pp. 316–322, 2005.
[5] C. Wang, J. Zhang, L. Wang, J. Pu, and X. Yuan, “Human iden-
tification using temporal information preserving gait template,”
IEEE transactions on pattern analysis and machine intelligence,
vol. 34, no. 11, pp. 2164–2176, 2011.
[6] H. Chao, Y . He, J. Zhang, and J. Feng, “Gaitset: Regarding gait
as a set for cross-view gait recognition,” in Proceedings of the
AAAI conference on artificial intelligence, vol. 33, no. 01, 2019,
pp. 8126–8133.
[7] C. Fan, Y . Peng, C. Cao, X. Liu, S. Hou, J. Chi, Y . Huang,
Q. Li, and Z. He, “Gaitpart: Temporal part-based model for
gait recognition,” in Proceedings of the IEEE/CVF conference
on computer vision and pattern recognition, 2020, pp. 14 225–
14 233.
[8] R. Liao, S. Yu, W. An, and Y . Huang, “A model-based gait
recognition method with body pose and human prior knowledge,”
Pattern Recognition, vol. 98, p. 107069, 2020.
[9] K. Lenac, D. Suˇ sanj, A. Ramaki´ c, and D. Pinˇ ci´ c, “Extending
appearance based gait recognition with depth data,” Applied
Sciences, vol. 9, no. 24, p. 5529, 2019.
[10] A. Ramaki´ c, D. Suˇ sanj, K. Lenac, and Z. Bundalo, “Depth-Based
Real-Time Gait Recognition,” Journal of Circuits, Systems and
Computers, p. 2050266, 2020.
[11] D. Pinˇ ci´ c, D. Suˇ sanj, and K. Lenac, “Gait recognition with self-
supervised learning of gait features based on vision transform-
ers,” Sensors, vol. 22, no. 19, p. 7140, 2022.
[12] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classi-
fication with deep convolutional neural networks,” Advances in
neural information processing systems, vol. 25, 2012.
[13] Z. Wu, Y . Huang, L. Wang, X. Wang, and T. Tan, “A compre-
hensive study on cross-view gait based human identification with
deep cnns,” IEEE transactions on pattern analysis and machine
intelligence, vol. 39, no. 2, pp. 209–226, 2016.
[14] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn,
X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold,
S. Gelly et al., “An image is worth 16x16 words: Transformers
for image recognition at scale,” arXiv preprint arXiv:2010.11929,
2020.
[15] K. Bashir, T. Xiang, and S. Gong, “Gait recognition using gait
entropy image,” 2009.
[16] M. Rokanujjaman, M. A. Hossain, M. R. Islam, and M. S. Islam,
“Effective part definition for gait identification using gait entropy
image,” in 2013 International Conference on Informatics, Elec-
tronics and Vision (ICIEV). IEEE, 2013, pp. 1–4.
[17] M. Jeevan, N. Jain, M. Hanmandlu, and G. Chetty, “Gait recog-
nition based on gait pal and pal entropy image,” in 2013 IEEE
International Conference on Image Processing. IEEE, 2013,
pp. 4195–4199.
[18] S. Yu, D. Tan, and T. Tan, “A framework for evaluating the effect
of view angle, clothing and carrying condition on gait recogni-
tion,” in 18th International Conference on Pattern Recognition
(ICPR’06), vol. 4. IEEE, 2006, pp. 441–444.
[19] R. Liao, W. An, Z. Li, and S. S. Bhattacharyya, “A novel
view synthesis approach based on view space covering for gait
recognition,” Neurocomputing, vol. 453, pp. 13–25, 2021.
[20] J. Kovaˇ c, V .
ˇ
Struc, and P. Peer, “Frame–based classification for
cross-speed gait recognition,” Multimedia Tools and Applica-
tions, vol. 78, pp. 5621–5643, 2019.
[21] C. Shannon and W. Weaver, “The mathemtiatical theory of
comnunication,” Urbana: University of Illinois Press, 1949.
[22] M. Masi, “A step beyond tsallis and r´ enyi entropies,” Physics
Letters A, vol. 338, no. 3-5, pp. 217–224, 2005.
[23] T. Maszczyk and W. Duch, “Comparison of shannon, renyi and
tsallis entropy used in decision trees,” in Artificial Intelligence
and Soft Computing–ICAISC 2008: 9th International Conference
Zakopane, Poland, June 22-26, 2008 Proceedings 9. Springer,
2008, pp. 643–651.
[24] C. F. L. Lima, F. M. de Assis, and C. P. de Souza, “An

52 DUMEN
ˇ
CI
´
C ET. AL.
empirical investigation of attribute selection techniques based
on shannon, r´ enyi and tsallis entropies for network intrusion
detection,” American Journal of Intelligent Systems, vol. 2, no. 5,
pp. 111–117, 2012.
[25] D. Suˇ sanj, V . Tuhtan, L. Lenac, G. Gulan, I. Koˇ zar, and
ˇ
Z. Jeriˇ cevi´ c, “Using entropy information measures for edge de-
tection in digital images,” in 2015 38th International Convention
on Information and Communication Technology, Electronics and
Microelectronics (MIPRO). IEEE, 2015, pp. 352–355.
[26] J.-H. Ou and Y . K. Ho, “Shannon, r´ enyi, tsallis entropies and
onicescu information energy for low-lying singly excited states
of helium,” Atoms, vol. 7, no. 3, p. 70, 2019.
[27] O. Olendski, “R´ enyi and tsallis entropies: three analytic exam-
ples,” European Journal of Physics, vol. 40, no. 2, p. 025402,
2019.
[28] A. R´ enyi, “On measures of entropy and information,” in Pro-
ceedings of the Fourth Berkeley Symposium on Mathematical
Statistics and Probability, Volume 1: Contributions to the Theory
of Statistics, vol. 4. University of California Press, 1961, pp.
547–562.
[29] C. Tsallis, “Possible generalization of boltzmann-gibbs statis-
tics,” Journal of statistical physics, vol. 52, pp. 479–487, 1988.
[30] M. Hofmann, S. Bachmann, and G. Rigoll, “2.5D gait biometrics
using the Depth Gradient Histogram Energy Image,” in IEEE
Fifth International Conference on Biometrics: Theory, Applica-
tions and Systems (BTAS 2012), Arlington, Virginia, Sep 2012,
pp. 399–403.
[31] M. Caron, H. Touvron, I. Misra, H. J´ egou, J. Mairal, P. Bo-
janowski, and A. Joulin, “Emerging properties in self-supervised
vision transformers,” in Proceedings of the IEEE/CVF Interna-
tional Conference on Computer Vision, 2021, pp. 9650–9660.
[32] H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles,
and H. J´ egou, “Training data-efficient image transformers &
distillation through attention,” in International Conference on
Machine Learning. PMLR, 2021, pp. 10 347–10 357.
Stella Dumenˇ ci´ c is a researcher and PhD student at the University
of Rijeka, Faculty of Engineering. She received her Bachelor’s degree
in Computing in 2021 and her Master’s degree in Computing in 2023
from the University of Rijeka, Faculty of Engineering. Her research
interests are in the fields of machine learning and mathematical
optimizations.
Dr. Domagoj Pinˇ ci´ c received the B.S., M.S. and Ph.D. degrees in
Computer Engineering and Science from the Faculty of Engineering,
University of Rijeka, in 2015, 2018 and 2022, respectively. His
research interests include computer vision, machine learning, and
image processing.
Dr. Diego Suˇ sanj is a Postdoctoral Researcher at the University of
Rijeka, Faculty of Engineering. He received his Bachelor’s degree
in Computer Engineering in 2013, his Master’s degree in Computer
Engineering in 2015 and his PhD in Computer Science in 2021
from the University of Rijeka, Faculty of Engineering. His research
interests are in the fields of computer vision, machine learning, image
processing and embedded systems.
Dr.
ˇ
Ziga Emerˇ siˇ c graduated from the Faculty of Computer and
Information Science, University of Ljubljana, where he also holds
a position of assistant professor. His research interests include deep
learning and biometrics which intersect with his teaching areas. He
has co-authored more than 40 research papers and received multiple
awards for teaching and research, including the European Association
for Biometrics Award Max Snijder 2021.