2-IJCI Vol. 3 No. 6-June 2024-Paper1-Dr. Saadi

Alsaif Dataset Creation and Evaluation for
Gymnastics Movement
Khalil I. Alsaif
Department of Computer Techniques Engineering, Al-Hadbaa University
College, Mosul, Iraq
Khalil_alsaif@hcu.edu.iq
Ahmed Saadi Abdullah

Assistant professor, College of Computer Science and Mathematics, Mosul
University, Mosul, Iraq
ahmedalbasha@tu.edu.iq
Abstract
Computer vision systems play a major role in many areas of life, as reliance has clearly
begun on these systems in the medical field, where they help doctors diagnose diseases
correctly. In addition, they are relied upon in the sports field, where they can help
referees in giving correct arbitration decisions. The accuracy of computer vision
systems has increased with the emergence of deep learning techniques, which require
a large amount of data on which a deep learning model can be trained. Therefore, if
any deep learning model is relied upon to detect and distinguish any movement or any
entity within images or video clips, it must first Training this model on a set of images
of these movements or objects to be discovered and distinguished. From this
standpoint, this article presents the construction of a database for one of the important
games, the gymnastics game, which is popular in many countries. A database has been
built for the gymnastics game, which can be used in the application. Deep learning
models were developed in order to help coaches and referees in this game, and for
these movements there is no database available on the Internet. A group of video clips
9
International Journal of Computers and Informatics, London Vol (3), No (6), 2024
https://doi.org/10.59992/IJCI.2024.v3n6p1 E-ISSN 2976-9361
spread on the Internet was relied upon to build this base, where ten basic movements
in the game of gymnastics were distinguished, and one of the models was evaluated.
Deep learning on these images (yolo7).
Keywords: Gymnastics Sports, Creation Dataset, Computer vision, Deep Learing.
Introduction
The player must possess a variety of skills in order to participate in the gymnastics
game. These skills include the ability to maintain balance, flexibility, and strength.
The reason for this is that the majority of the actions in this game are dependent on
these characteristics. These characteristics include those that depend on strength, such
as stability movements, as well as certain movements that require a high amount of
flexibility, particularly in certain postures. In this game, the arbitration procedure and
point distribution for backward somersaults are very dependent on the player's ability
to execute these movements in the exact sequence that is specified [1].
The analysis of the gymnast's motions is a highly significant topic. Analyzing these
motions can be advantageous as it can elucidate the player's shortcomings and prevent
possibly injurious movements. The reason for this is that conventional approaches
depend on the deployment of sensors positioned at different areas on the player's body.
These sensors collect data, which is then used to calculate values and quantify the
player's performance [2].
Among the subfields of computer science is the field of computer vision. In the present
moment, it has been utilized in a wide range of applications and fields, including
engineering, industry, and medicine. Complex mechanisms were utilized by the
factories in the past in order to monitor the products. When monitoring the product,
the worker must exercise extreme caution. Some of the products that have been passed
over might not be good if the worker is ignored or if their expertise is insufficient. if
10
this occurs, the process will be accurate, but at the same time it will exhaust the
workers, which will result in an increased demand for additional workers [3].
When it comes to any field, whether it be medicine, sports, or any other field, the
primary objective of constructing a computer vision system is to acquire an image.
When that is done, the component of his image can be analyzed, which means studying
its content. an example of this would be the identification of the movement of an object
or a particular color in an image through tracking it within that image, particularly in
the context of medical diagnosis [2,3]. to be more specific, the process of constructing
a computer vision system is a simulation of the human vision system. The difference
between them lies in the fact that the human eye is used in the human vision system,
whereas digital cameras are used in the computer vision system [2-4].
As can be seen in figure (1), the algorithms that are utilized in the process of examining
the relationship between the components are analogous to the method in which the
human mind operates. Human eyesight versus computer vision is depicted in Figure
1. The use of computer vision systems and deep learning has become increasingly
prevalent in a variety of sectors, including sports, in order to achieve the most
favorable outcomes. Computer vision and deep learning techniques are utilized in the
field of sports to assess the movement of players inside stadiums. This analysis is
performed in the context of the sports industry. Through the use of these tactics,
coaches are able to select the playing strategy as well as the best players who are
capable of putting this strategy into action by studying the findings gained from
analyzing the performance of the player.2[4].
The Application of Computer Vision in Olympic Sports Among the numerous fields
that make use of computer vision is the sporting world. There are a number of
computer vision systems that can be deployed in this respect for the purpose of
determining the location of players or the ball, as well as tracking their journey inside
11
the field. In addition, coaches can reap the benefits of such methods when it comes to
the process of studying the strategy employed by their opponent. Due to the fact that
these systems could be of assistance to the referee in order to rectify a decision that
might be made as a result of insufficient view, this decision will be erroneous. A
scenario like this demonstrates the need of utilizing computer vision technology [5].
Figure (1): Human vision vs. Computer Vision

Deep learning:
This field has brought about fundamental changes in a number of different areas,
including the deep learning field, which was created by the neural network (NN) sub-
field on it. Deep learning is a specialized branch of the machine learning field that was
inspired in the 1950s by models of computation and human cognition. Deep learning
12
represents a small branch in the artificial intelligence field. Since it was first
introduced, deep learning has been responsible for even more significant disruptions,
demonstrating resounding success in every field of application [6-8]
Deep learning can be either a hierarchical learning model or a deep construction
method. Learning refers to a method that is used to estimate the parameters of the
model in order to design an algorithm that is designed to perform a specific task as
well as possible. On the other hand, deep learning has numerous layers between input-
output layers that allow for many non-linear processing units for information in several
stages of a hierarchical structure. This structure is used for the purpose of feature
learning. Learning approaches that are based on data representation are referred to as
representation the learning, and some recent work has described deep learning (DL)
as a learning model that is comprehensive and capable of delivering solutions to a
wide variety of issues in a variety of scientific domains [9].
There is a collection of techniques known as representational learning, which involves
providing a machine with unprocessed data in order to discover the representations
that are required for the categorization and discovery process. Deep learning
techniques are representational learning modes that consist of several levels of
representations produced from simple non-linear units. These representations are
transformed from lower to higher levels in a hierarchical fashion, and with sufficient
transformation, they are able to learn extremely complex functions. This is the point
of strength when compared to the traditional approaches that are used in machine
learning (Deep learning is based on data-driven learning algorithms, which are used
to generate feature layers. These feature layers are represented by various levels in a
hierarchy [10-12].
The Popular Concept of Deep Learning
The intelligent system (AI) can learn and improve experiences based on the experience
13
gained through data known as machine learning (ML), where machine learning
algorithms require that the input data be represented appropriately for the purpose of
giving correct predictions. Each representation for the project characteristics which
enables a system to make decisions is known as a feature, as learning by representation
(RL) helps machine learning algorithms not only learn the feature map but can also
help the representation itself. But the problem of learning by representations is that it
does not help in solving feature extraction, which represents essential abstract
properties that bring the system to critical decisions in prediction in realism
applications. Deep learning (DL) works on studying these difficulties and building
complex representations based on simple representations and having multi-layered
abstractions [13-15]. The algorithm allows models consisting of several processing
layers to work on and learn to represent that data using abstractions from several
layers. Figure 2. shows the relationship between the field of deep learning, teaching
representations, the discipline of machine learning, and artificial intelligence's last
field [16].
Figure 2: The relationship between fields, AI, ML, RL, and DL
14
Gymnastics Sprot
Gymnastics is a sport that requires strength, flexibility, balance, coordination, and
physical fitness. It is a comprehensive sport that includes performing complex and
varied movements on different tools or directly on the ground. Here is an explanation
of the most important types of gymnastics [17-19]:
Artistic Gymnastics: This category includes performing exercises on equipment such
as vault, uneven bars, balance beam, and floor exercises. This category relies on
strength, flexibility, balance and coordination to execute artistic movements with the
utmost precision and mastery.
Rhythmic Gymnastics: This category relies on performing exercises coordinated
with music using hand tools such as rope, ring, ball, plates, and ribbons. Players
combine power, precise control and elegance in executing movements.
Trampoline Gymnastics: This category includes performing acrobatic movements
and high jumps on a trampoline. Skills in this category require strength, air control,
and coordination to execute movements accurately and at the highest possible height.
Aerobic Gymnastics: This category focuses on performing exercises that contain
repetitive aerobic elements, such as jumping, spinning, and strong movements, in
coordination with music. This class requires strong fitness, flexibility and physical
coordination.
Gymnastics requires hard work and regular training to develop the necessary skills. It
is considered an exciting sport and is very popular all over the world, whether for
participation in sports competitions or as a recreational activity for children and adults
alike.
15
Dataset Creation
In this paragraph, we will discuss the method by which a dataset was built for the
movements of the gymnastics player, which is not available on the Internet. Therefore,
this is considered the first database in this field that will be available on the Internet.
It was built to study 10 basic movements in the game of gymnastics, which are the
movements of strength and stability in the throat apparatus, which It is intended for
men because these movements require a lot of strength to perform, and therefore they
will be intended for men The next step will explain the process of building it in detail,
in addition to its website through which one can access and deal with it.
Data Collection
Hence, it is apparent that this game lacks popularity in numerous regions. The
variation in popularity of these games across different countries is attributed to the
differing levels of infrastructure required for this genre of game. Consequently, these
rooms are not universally accessible due to a scarcity of certified gyms in various
countries. In Iraq, we utilized a compilation of video clips and images that were widely
circulated on the internet. Additionally, we incorporated a small assortment of pictures
obtained from the Education for Physical Sciences department at Tikrit University.
After the event, we depended on high-quality video clips. This compilation of video
clips was first converted into a series of individual frames, which were subsequently
converted into a series of photographs. Figure 4 displays an example of one of these
images.
16
Figure (4): ALsaif Dataset
Data Augmentation
Deep learning models require a substantial quantity of photos during the training
process, and the accuracy of the outputs generated by the model is directly influenced
by the amount of images utilized in training. Consequently, it is preferable to train the
model using a substantial volume of data in order to achieve extremely precise
outcomes and so avoid inaccuracies. It is essential for the model to be trained on a
large number of images. However, it is preferable for these images to have diversity
in terms of lighting conditions and viewing angles of the objects. Therefore, it is ideal
to avoid obtaining these effects. It is more advantageous to employ software that can
apply a set of effects to these photographs, including modifying the composition and
rotating them from multiple perspectives.
Data Annotations
At this point, annotations will be made on the images of the gymnast player's moves
that were captured on the static ring apparatus in the gymnastics game contained in
the dataset. In addition, a thorough investigation was conducted on the ten most
17
significant movements in this game. To enable the deep learning model to distinguish
these motions, it is necessary to accurately identify the images of these movements.
During the testing phase, we utilized the roboflow application to complete the task of
annotating the images with comments. Figure 5 showcases a compilation of photos
that have been accompanied by remarks.
Figure (5): ALsaif Dataset Annotation

The annotated technique will result in squares that are enclosed by images of the
gymnast's movements and contain a compilation of information. The information will
be contained within these squares. The framework of the deep learning model (yolov7)
can collect five distinct bits of information. Aside from the image's locations inside
the original image, the primary piece of information is the image's classification.
Furthermore, the dimensions of the image depicting the gymnast's movements, which
will serve as the training data for the deep learning model, are also provided in this
information. Figure 6 displays an instance of the format that is stored in text format.
18
Figure (6): Label Alsaif Dataset
Dataset Splitting
After the data has been collected and a set of effects have been added to the images,
the data is now divided into groups for training and testing. The following algorithm
shows the method of dividing the data.
19
Evaluation Alsaif Dataset
The deep learning model employed in this thesis was assessed on different dataset
partitions to identify the most accurate split of the data. The dataset was split into two
halves, with 70% of it allocated for training and the remaining 30% for testing.
Furthermore, it was expected that a quarter of the data would be allocated for testing,
whereas three-quarters of the data would be allocated for the testing process. During
the training phase, the algorithm showed that the best division happened when 100%
of the data was used, which corresponded to eighty percent. Based on the performance
of the deep learning model provided in this thesis, Figure 7 illustrates an evaluation of
the model's performance on different partitions of the dataset.
Figure (7): Evaluation Alsaif Dataset
20
Conclusion
Relying on computer vision systems in the sports field is very important, especially in
the arbitration process, in order to obtain high accuracy in arbitration decisions instead
of relying on the human vision system, which may be subject to error as a result of
poor vision for the referees, in addition to the fact that referees sometimes have little
experience in Some arbitration cases, or the referee may be sympathetic to one of the
teams. Providing databases for games that are not widely spread, such as football or
basketball, will open new horizons for researchers to apply models of deep learning to
these data. This database contains approximately 2,200 divided images. There are 10
types of movements for the gymnast on the throat apparatus, and the database is
available on the Rap Flow website and at the following link:
https://app.roboflow.com/alsaifforgymnastic-edhbc/alsaif_for_gymnastic/2
References
1. Ahmed Saadi Abdullah, Khalil I. Alsaif, (2023),” Still Rings Movements Recognition in
Gymnastics Sport Based on Deep Learning”, Wasit Journal for Pure Sciences, Vol. (2) No. (1).
2. Shae Donham-Foutch, (2007)” Teaching skills and health-related _tness through a preservice
gymnastics program”, Journal of Physical Education, Recreation & Dance, 78(5).
3. Ahmed S. Abdullah1, Khalil Ibrahim alsaif , (2023)” Computer vision systems and deep learning
for the recognition of athlete's movement: A review article :, Tikrit Journal of Pure Science
(2023) 28 (6), https://doi.org/10.25130/tjps.v28i6.1600
4. Athanasios Voulodimos Nikolaos Doulamis, Anastasios Doulamis, and Eftychios
Protopapadakis, (2018)” Deep Learning for Computer Vision: A Brief Review”, Computational
Intelligence and Neuroscience journal, 2018.
5. M. Waleed, A. S. Abdullah and S. R. Ahmed, “Classification of Vegetative pests for cucumber
plants using artificial neural networks,” 2020 3rd International Conference on Engineering
Technology and its Applications (IICETA),
21
6. A. S. Abdullah and K. I. Alsaif, (2023)"Recognition and Evaluation of Stability Movements in
Gymnastics Based on Deep Learning," 2023 Al-Sadiq International Conference on
Communication and Information Technology (AICCIT), Al-Muthana, Iraq, 2023, pp. 267-271,
doi: 10.1109/AICCIT57614.2023.10218071.
7. Majid Hamid Ali and et, (2019),” Evaluation of the Performance of Kirsch and Robinson
Methods on Different Color Spaces”, DOI 10.5013/IJSSST.a.20.04.05.
8. majida ali abd and et, (2019),” Traffic signs recognition using cuckoo search algorithm and
Curvelet transform with image processing methods”, Journal of AL- Qadisiyah for computer
science and mathematics.
9. Naik, Banoth Thulasya, Mohammad Farukh Hashmi, and Neeraj Dhanraj Bokde. (2022). "A
Comprehensive Review of Computer Vision in Sports: Open Issues, Future Trends and Research
Directions" Applied Sciences Journal, vol 12, no. 9, pp. 1-49,
https://doi.org/10.3390/app12094429.
10. Lijin Zhu, (2021). "Computer Vision-Driven Evaluation System for Assisted Decision-Making
in Sports Training", Wireless Communications and Mobile Computing Journal vol 2021, no. 7,
pp. 1–7, https://doi.org/10.1155/2021/1865538
11. Kim Hébert- Losier, Peter Lamb and C. Martyn Beaven. (2022). "Biomechanical determinants
of placekicking success in professional Rugby Union players ". Sports Biomechanics vol 21
no.7, pages 861-876, https://doi.org/10.1080/14763141.2020.1728368.
12. Rana Abdulrahman Lateef and Ayad Rodhan Abbas, (2022). "Human Activity Recognition
using Smartwatch and Smartphone: A Review on Methods, Applications, and Challenges", Iraqi
Journal of Science, vol. 63, No. 1, pp: 363-379. https://doi.org/10.24996/ijs.2022.63.1.34
13. Hang Pan and Jianwei Li and Huidong Wang and Kaiyu Zhan, (2021). "Biomechanical analysis
of shooting performance for basketball players based on Computer Vision", Journal of Physics:
Conference Series vol.2024 no.1 pp: https://doi.org/1-11. 10.1088/1742-6596/2024/1/012016
14. Z. Rybchak, O. Basystiuk, (2017). "Analysis of computer vision and image analysis technics ",
Econtechmod an International Quarterly Journal vol. 6. no. 2. pp: 79–84.
15. S. Manish., V. Bhagat and R. Pramila, (2021), "Prediction of Football Players Performance
22
using Machine Learning and Deep Learning Algorithms," 2nd International Conference for
Emerging Technology (INCET), pp. 1-5, https://doi.org/10.1109/INCET51464.2021.9456424.
16. Chen Huang, Lei Jiang, (2021). "Data monitoring and sports injury prediction model based on
embedded system and machine learning algorithm.", Microprocessors and Microsystems journal
vol.81. https://doi.org/10.1016/j.micpro.2020.103654.
17. Roman Farana, Timothy Exell, Gerda Strutzenberger & Gareth Irwin, (2022). "Technique
selection in young female gymnasts: Elbow and wrist joint loading during the cartwheel and
round-off", European Journal of Sport Science vol.18, pp:423-430.
https://doi.org/10.1080/17461391.2018.1424941
18. Jamal Sakran Hamza, (2018), "Historical development of the apparatus and the code men
Artistic gymnastics until (2017) and its impact on the level of motor difficulties", Ph. D thesis,
University of Baghdad.
19. Rizzato, Alex, Antonio Paoli, and Giuseppe Marcolin. (2021). "Different Gymnastic Balls
Affect Postural Balance Rather Than Core-Muscle Activation: A Preliminary Study" Applied
Sciences 11, no. 3. https://doi.org/10.3390/app11031337
23

2-IJCI Vol. 3 No. 6-June 2024-Paper1-Dr. Saadi

Uploaded by

Copyright:

Available Formats

2-IJCI Vol. 3 No. 6-June 2024-Paper1-Dr. Saadi

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

2-IJCI Vol. 3 No. 6-June 2024-Paper1-Dr. Saadi

Uploaded by

Copyright:

Available Formats

Alsaif Dataset Creation and Evaluation for

Ahmed Saadi Abdullah

Figure (1): Human vision vs. Computer Vision

Figure 2: The relationship between fields, AI, ML, RL, and DL

Figure (5): ALsaif Dataset Annotation

Figure (7): Evaluation Alsaif Dataset

You might also like