1,721,005 research outputs found
Multimodal AI-Based Summarization and Storytelling for Soccer on Social Media
No full textSoccer is not only a beloved global sport but also a significant industry that generates substantial revenue through sponsorship, merchandise, broadcast rights, and ticket sales. In today’s digital age, the way fans consume soccer content is evolving. A growing number of fans now prefer to watch or read game highlights rather than full matches, a shift driven by busy lifestyles and a preference for brief, digestible content formats. This change is facilitated by digital platforms that cater to the demand for quick and engaging soccer insights.
In this changing landscape, the role of Artificial Intelligence (AI) is becoming increasingly crucial. AI technology has the potential to streamline the production of game highlights by automatically detecting and compiling key moments such as goals and significant plays. This not only saves time and reduces errors but also enhances the consistency of media output, aligning production with consumer preferences for quick and accessible content.
Building on this premise, this thesis aims to develop an AI-driven method specifically designed to automatically generate multi-tweets thread summarizing soccer goals. This innovative approach allows fans to visualise the essence of game highlights without needing to watch the full event. It also offers soccer clubs a cost-effective strategy to engage with fans on social media platforms like Twitter, potentially boosting long-term revenue in this evolving business era.
To achieve these objective, SoccerSum leverages a multimodal AI-based system, developed and extensively tested through objective and two rounds of subjective evaluations. SoccerSum combines video, audio, and text modalities to generate concise, engaging summaries for goal events specifically designed for Twitter. Evaluation results demonstrate that SoccerSum excels in delivering accurate name entities, integrating pitch elements effectively, and using language that resonates with users.
Keywords: AI, Machine Learning (ML), Social Media, Content Generation, Generative A
SmartCrop: AI-Based Cropping of Sports Videos
No full textSports multimedia is among the most prominent types of content distributed across
social media today, necessitating the retargeting of videos to diverse aspect ratios for
appropriate representation on different platforms. SmartCrop is an automated video
cropping pipeline designed to curate content tailored to custom aspect ratios suitable
for various social media platforms. The system utilizes a Point of Interest (POI)
tracking mechanism, with the soccer ball or ice hockey puck serving as the primary
POI. Scene detection is achieved through TransNetV2 (a machine learning model) and
PySceneDetect (a Python library), while a You Only Look Once (YOLO)v8-medium
model, fine-tuned on custom soccer and ice hockey datasets, detects the POIs.
Inaccurate detections are filtered through outlier detection methods, and
interpolation or smoothing modules are applied when the POI is not visible, specific
to either soccer or ice hockey. Objective evaluations of each module’s performance
within both the SmartCrop-S and SmartCrop-H pipelines have been conducted,
validating the proposed architecture in terms of accuracy, efficiency, precision, and
error metrics such as RMSE and MAE. These evaluations confirm that the system
meets high standards for performance and is effectively adapted to the dynamic
requirements of sports video analysis. For the SmartCrop-S pipeline, a crowdsourced
subjective user study assessing alternative cropping approaches from 16:9 to 1:1 and
9:16 aspect ratios confirms that the proposed approach significantly enhances the
end-user Quality of Experience (QoE). For the SmartCrop-H pipeline, three distinct
subjective user studies were conducted: the first to determine the optimal alpha value
for the smoothing module, the second to show that the SmartCrop output using the
full functionality of the SmartCrop-H pipeline performed better than other
alternatives, and the third, designed for competitor analysis, compared SmartCrop-H
with professional video editing tools. This last study demonstrated that SmartCrop-H
performs on par with, or even surpasses, professional tools in terms of output quality
AI-based clipping of booking events in soccer
Full text linkManual clipping is currently the gold standard for extracting highlight clips from soccer games. However, it is a costly, tedious, and time-consuming task that is impractical and unfeasible for, at least, lower-league games with limited resources. Today, the manual method is to use a preset time interval, trimming away undesired video frames. To address this issue, this thesis aims to automate the generation of highlight clips for booking events. In our pipeline, we will implement logo detection, scene boundary detection, and multimedia processing. We will also do a statistical analysis of current highlight clips, and perform a subjective evaluation. Full games are used as input, where detection modules will locate possible timestamps to produce an intruguing highlight clip. Through experimentation and results from state-of-the-art research, we will use neural network architectures and different datasets to suggest two models that can automatically detect appropriate timestamps for extracting booking events. These models are evaluated both qualitatively and quantitatively, demonstrating high accuracy in detecting logo and scene transitions and generating viewer-friendly highlight clips. When looking at state-of-the-art research and the results in the thesis, the conclusion is that automating the soccer video clipping process has significant potential.publishedVersio
Smartphone-Based Crowdsourcing for Mobile Network Benchmarking
No full textCrowdsourcing is a neologism coined as a portmanteau of crowd and outsourcing. The crowdsourcing paradigm is introduced
to a plethora of disciplines, but its potential applications in telecommunications, in particular for the
performance evaluation and benchmarking of mobile networks has not yet been fully explored. This
research is specifically targeting at establishing such a framework.Crowdsourcing is a neologism coined as a portmanteau of crowd and outsourcing. The crowdsourcing paradigm is introduced
to a plethora of disciplines, but its potential applications in telecommunications, in particular for the
performance evaluation and benchmarking of mobile networks has not yet been fully explored. This
research is specifically targeting at establishing such a framework
Soccer athlete performance prediction using time series analysis
No full textRegardless of the sport you prefer, your favorite athlete has almost certainly
disappointed you at some point. Did you jump to a conclusion and
dismissed it as "not their day"? Or, did you consider the underlying causes
for their poor performance on that particular day? Under-performance can
have big consequences in team sports such as soccer and affect the entire
team dynamic. Basal needs like sleep quality and wellness parameters such
as mood, fatigue and muscle soreness can affect an athlete’s performance.
In this context, the practice of using wearable sensor devices to quantify
athlete health and performance is gaining popularity in sports science. This
thesis aims to predict how ready a soccer athlete is to train or play a match
based on the subjectively reported wellness parameter readiness to play,
collected by the PMSys athlete performance monitoring system [34, 33, 17].
Even though women’s soccer is receiving increasingly more attention, with
a recent record in game day attendance marking over 90.000 spectators [50],
the vast majority of soccer studies are conducted on male athletes. In this
sense, we explore a relatively new domain using the PMSys dataset, which
is from two Norwegian elite female soccer clubs over the period of 2020
and 2021. We predict readiness by utilizing the Long short-term memory
(LSTM) method and the Tsai [45] state-of-the-art deep learning library. We
develop a framework that is able to handle univariate multistep time series
prediction and easily allows for further development. The experimental
results show that it is possible to train a Machine Learning (ML) model
on a team and predict a single player’s readiness, detecting detect peaks
closely to actual values. It is possible to use the previous week to predict
the upcoming day, or even the upcoming week, as the model does not
require much data to get started. The model works well on data from the
entire team for a shorter period than a larger set of data for a longer period,
which allows the teams to quickly start using the system with existing data.
Hyperparameters are easily configurable and can be changed as required to
optimize the model. Our results can be used for evidence based decisions,
such as benching the team star so she doesn’t get injured for the rest of the
season. As a first milestone, this framework will be incorporated in PMSys
and used in the Norwegian the elite female soccer league, Toppserien, but
the overall approach can be part of a standardized athlete performance
monitoring system that is globally used by athletes in all sports
Soccer athlete performance prediction using time series analysis
Full text linkRegardless of the sport you prefer, your favorite athlete has almost certainly
disappointed you at some point. Did you jump to a conclusion and
dismissed it as "not their day"? Or, did you consider the underlying causes
for their poor performance on that particular day? Under-performance can
have big consequences in team sports such as soccer and affect the entire
team dynamic. Basal needs like sleep quality and wellness parameters such
as mood, fatigue and muscle soreness can affect an athlete’s performance.
In this context, the practice of using wearable sensor devices to quantify
athlete health and performance is gaining popularity in sports science. This
thesis aims to predict how ready a soccer athlete is to train or play a match
based on the subjectively reported wellness parameter readiness to play,
collected by the PMSys athlete performance monitoring system [34, 33, 17].
Even though women’s soccer is receiving increasingly more attention, with
a recent record in game day attendance marking over 90.000 spectators [50],
the vast majority of soccer studies are conducted on male athletes. In this
sense, we explore a relatively new domain using the PMSys dataset, which
is from two Norwegian elite female soccer clubs over the period of 2020
and 2021. We predict readiness by utilizing the Long short-term memory
(LSTM) method and the Tsai [45] state-of-the-art deep learning library. We
develop a framework that is able to handle univariate multistep time series
prediction and easily allows for further development. The experimental
results show that it is possible to train a Machine Learning (ML) model
on a team and predict a single player’s readiness, detecting detect peaks
closely to actual values. It is possible to use the previous week to predict
the upcoming day, or even the upcoming week, as the model does not
require much data to get started. The model works well on data from the
entire team for a shorter period than a larger set of data for a longer period,
which allows the teams to quickly start using the system with existing data.
Hyperparameters are easily configurable and can be changed as required to
optimize the model. Our results can be used for evidence based decisions,
such as benching the team star so she doesn’t get injured for the rest of the
season. As a first milestone, this framework will be incorporated in PMSys
and used in the Norwegian the elite female soccer league, Toppserien, but
the overall approach can be part of a standardized athlete performance
monitoring system that is globally used by athletes in all sports.publishedVersio
An Open Dataset of Operational Mobile Networks
Full text linkMobile networks have become ubiquitous and the primary meansto access the Internet, and the traffic they generate has rapidlyincreased over the last years. The technology and service diversityin mobile networks call for extensive and accurate measurementsto ensure the proper functioning of the networks and rapidly spotimpairments. However, the measurement of mobile networks iscomplicated by their scale, and, thus, expensive, especially due tothe diversity of deployments, technologies, and web services. Inthis paper, we present and provide access to the largest open in-ternational mobile network dataset collected using the MONROEplatform spanning six countries, 27 mobile network operators, and120 measurement nodes. We use them to run measurements tar-geting several web services from January 2018 to December 2019,collecting millions of TCP and UDP flows using these commercialmobile networks. We illustrate the data collection platforms and de-scribe some of the main experiments. Besides a high-level overviewof the dataset, we provide two practical use cases. First, we showhow our data can be used as a proxy for web service performance.Second, we study the content delivery infrastructure of Facebook
Cloud Operations to Support UX and Accessibility for Crowdsourced Online Survey Framework Deployment
Full text linkThis study investigates different cloud service providers (CSPs) and their deployment platforms in terms of user friendliness and experience. With the help of Docker virtualization, allowing to create an isolated environment, the application deployment process can be replicated on different platforms and prevent technical issues linked to the underlying technical configurations of a CSP. It further allows for non-technical users to choose their CSP based on requirements other than the necessary technical aspects and deploy their applications. To ensure a higher level of user experience (UX) and accessibility for application deployment, we investigate ways of optimizing cloud operations. Docker is used to create a container image along with Docker Compose to define multi-container applications. The Docker Compose files are created as base files for further development with a production and development environment. Using these files, an evaluation is conducted on the features, services, and deployment processes of different CSPs. We conducted a deployment test using both a web-based GUI and a terminal for the CSP, which resulted in a consistent and comparable deployment. As discovered through testing, using the developed Dockerfiles prevent certain issues of the technical infrastructures of CSPs. As part of the thesis, a subjective user study is conducted consisting of a pre-questionnaire, a deployment guide and a post-questionnaire involving System Usability Scale (SUS). This user study uses the Huldra application for deployment on a CSP named Render. The pre-questionnaire responses regarding familiarity with computer science, application deployment, and Docker did not appear to significantly impact the deployment time. We also analyzed the relationship between the variables in terms of time and difficulty level of each deployment step were we found few significant trends. We also found low ratings of difficulty throughout the user study and a high mean score in participants responses with respect to user experience.publishedVersio
Analyzing and Benchmarking the Performance of Different Cloud Services for Agile App Deployment
Full text linkCloud computing has become a popular method for organizations to deploy and host websites. The cloud offers several innovative solutions and services for these tasks, but they also bring new challenges. Selecting the appropriate cloud platform and deployment method for a specific use case can be challenging due to the wide variety of options available in the market. The selection is also usually based on the specific requirements and technologies of the use case. This thesis aims to assist in the process of selecting a cloud platform and provide benchmarks and analysis of several cloud service providers (CSPs) and deployment methods. Two deployment methods are presented. The first method involves automatic deployment on CSPs through GitHub integration, while the second method utilizes a manual Dockerized approach with a virtual machine instance deployed in OpenStack. Analysis and benchmarks are conducted to obtain metrics based on resource usage and performance. The results obtained will be further analyzed to decide upon the CSP with the optimal performance. Huldra will be used as a test case for the deployment and hosting project. This is a React based survey framework developed by SimulaMet in Oslo, Norway. SimulaMet is currently exploring new approaches and services for deploying and hosting a Huldra survey in the cloud, which makes it a relevant use case. The thesis conclusion includes the approaches that can be applied to benchmark and analyze cloud services and deployment methods and assist in the selection of an alternative. The findings of the study are further presented which led to the conclusion that one CSP stood out as the optimal choice among the alternatives included.publishedVersio
SoccerRAG: Multimodal Soccer Information Retrieval via Natural Queries
Full text linkThe rapid growth and increasing availability of sports multimedia and metadata demands advanced information retrieval systems capable of efficiently processing vast amounts of multimodal data. This paper introduces SoccerRAG, an innovative framework that leverages Retrieval Augmented Generation (RAG), SQL agents, and Large Language Models (LLMs) to extract soccer-related information through natural language queries. By utilizing a multimodal dataset, SoccerRAG enables database querying, automatic data validation, and enhanced user interaction, making sports archives more accessible. Our evaluations demonstrate that SoccerRAG effectively handles complex queries and improves the accuracy of traditional SQL query systems. The results highlight the potential of RAG, agents, and LLMs in sports analytics, paving the way for future innovations in the accessibility and real-time processing of sports data
- …
