e-Repositori UPF :: Grau en Enginyeria de Sistemes Audiovisuals. Treballs de fi de grau

Recent Submissions

From data to understanding: music similarity through human perception, objective metrics and legal perspectives
(2025) Muñoz Melo, Lena
The rise of Artificial Intelligence (AI) in music offers creative potential but also raises complex challenges, especially in assessing originality and possible copyright infringement by generative models. The Music Replication Assessment (MiRA) tool was designed to identify potential data replication using audio-based similarity metrics. However, its alignment with human perception and legal standards has not been demonstrated. To explore this, we conducted a perceptual experiment comparing human judgments, MiRA classifications, and legal rulings on melodic similarity. Participants completed two tasks: (1) rating similarity on a scale, and (2) a forced choice to judge whether a fragment was a copy. Results revealed discrepancies among the three perspectives. Legal rulings aligned more with the forced choice task, while MiRA matched direct scaling, particularly in ambiguous cases.
Automatic trimming and scene matching in football videos
(2025) Soler Gonzàlez , Oriol
Football has established itself as one of the world’s most popular sports, and with today’s media consumption trends, there is increasing demand for accessible and targeted video content. In this thesis we present a first step toward developing an automatic video summarization tool for football matches. We create a tool to facilitate the creation of annotated datasets, essential for training and evaluating summarization models. The annotation process involves two steps: segmenting the summary into a set of shots and mapping each shot to its corresponding segment in the full game footage. To address these challenges, we explores state-of-the-art models for shot segmentation and scene matching. We propose a shot detection stage using TransNetV2, while in the matching stage we propose a backbone comparison using both 2D and 3D feature descriptors. The dataset we work with includes games from three major European leagues —Serie A, La Liga, and Ligue 1 — allowing for evaluation across diverse annotation styles and game formats.
Plan de empresa de SportsMate, aplicación que facilita la búsqueda de compañeros para actividades relacionadas con el ocio
(2025) Casado Chantrero , Jesmina
In today's increasingly digitalized society, the mobile device has become the primary tool for communication and information for a significant portion of the population. However, this trend has also led to the need for secure tools that facilitate social interaction among individuals with similar interests. SportsMate emerges as a solution to connect users with similar sports interests or hobbies. This document analyzes the potential of such a tool, as well as its technological and financial requirements. Furthermore, it addresses the development of appropriate functionalities with a contemporary design, adapted to the marketing needs of the target audience. The SportsMate application is considered to offer a unique opportunity to promote sociability and well-being among users, opening new possibilities for connection and community building in the sports and recreational field.
Pla de negoci de Travelize, una aplicació d’organització de viatges
(2025) Peña Salat, Xavier
La creixent demanda del sector turístic fa que cada vegada hi hagi més factors que influeixen en l’organització i la gestió dels viatges. Tot i l’existència de moltes aplicacions enfocades a aquest àmbit, realment hi ha poques alternatives que permetin organitzar els viatges de forma clara i ordenada des d’una mateixa plataforma. Per tant, sorgeix doncs la necessitat de crear Travelize. Feta a Catalunya i per a Catalunya, es tracta d’una aplicació mòbil d’organització de viatges plantejada amb l’objectiu de centralitzar l’organització dels viatges de les persones en una sola plataforma, amb una interfície totalment intuïtiva i amb tots els mitjans necessaris per a millorar de forma considerable l’experiència de les persones viatgeres. A través d’aquest Treball de Fi de Grau es presenta el pla de negoci de Travelize. Així mateix, s’estudien i s’avaluen tots els factors que poden afectar l’entrada al mercat d’aquesta aplicació. A més, també es determina la seva viabilitat financera, tot amb l’objectiu principal d’assegurar la sostenibilitat i fiabilitat del model de negoci presentat en aquest document.
Integrating harmonic power spectrum into essentia: a comparative study for pitch detection focused on flute instruments
(2024) Tortosa Arnau, Júlia; Tutor: Jorge Marcos
Essentia1 is an audio library created by the Music Technology Group designed to provide advanced audio and music analysis tools. While it includes state-of-the-art pitch detection algorithms like PitchMelodia or PitchYin, there are some algorithms that are not implemented like the Harmonic Product Spectrum (HPS). The HPS algorithm consists in applying some downsampling in the audio spectra and multiplying them to enhance the detection of the fundamental frequency peak, since it will be the highest peak over the other harmonics on the resulting spectrum. This project aims to integrate the HPS into Essentia library, coupled with a comparative analysis against existing algorithms within Essentia. The algorithm evaluation is centered on flute instruments and similar monophonic variants, aligning with algorithms best suited for such instruments. The project begins with a detailed theoretical exploration of the HPS algorithm, including important implementation considerations. Afterwards, the implementation process in C++ within a local Essentia repository is detailed, followed by compilation with Python bindings to access the new algorithm. Finally, the evaluation phase, conducted in Python, makes use of a dataset comprising audio files and the corresponding ground-truth files. Pitch detection algorithms from Essentia are employed to extract frequencies from the audio data and then the data is aligned for comparison with algorithmic results. The evaluated metrics are voicing recall and false alarm, pitch and overall accuracy and they are measured with mir_eval library.
Implementation of a real-time 3D navigation visualization on the web
(2023) Del Corral Farrarós, Carolina
Traditional nautical autopilot systems require ships to have complex and expensive electronic components, making its use and access difficult to the private user with small motor crafts. Under the need to cover this niche market, Searebbel was born, a start-up focusing on developing a digital rudder that connects to a mobile application. The app aims to make autonomous navigation easier and more accessible by means of route planning based on user-specified waypoints on a 2D map. In parallel it provides a 3D visualization simulating the users boat sailing in real-time based on weather conditions, sun position, and location, and a social component where users can add information buoys and share their journey with friends. This thesis covers the implementation of the pre-MVP with an interactive 2D map and a real-time dynamic 3D visualization of the maritime scene. The application offers the rendering of a large ocean surface in real-time with physics, as well as the support of multiple weather conditions such as rain, clouds, fog and different sea states, making it a challenging task to achieve in the context of mobile devices. This project uses the React library as the container of the frontend application, implements the MERN stack for the backend connection, Mapbox for the 2D scenery, and React Three Fiber, an extension of the Three.js library, for the graphics. The result of this work is a web demo that fulfills all the requirements with outstanding performance while still delivering great visual results, carrying out a proof of concept that can easily be extended in the future.
Interactive climbing wall using computer vision and real-time projection mapping
(2024) Panadès i Closes, Oriol
Interactive climbing walls are a technology that has gained popularity in recent years along with the rise of indoor climbing. The current limitations in the existing systems have motivated this project to investigate if a general, simple, and robust solution can be achieved using widely available tools. The developed system uses a regular projector, an uncalibrated camera (such as the one in a mobile phone), and a custom-built application to create an interactive climbing wall experience with real-time projection mapping. The implemented solution incorporates computer vision techniques such as image pre-processing, object detection and segmentation, feature matching, au- tomatic homography computation and ranking, as well as deep learning models for human pose estimation. For the application, a custom GUI has been created to facilitate easy operation by the user. The result is an interactive climbing wall that can be setup with minimal equipment, configured in less than 10 minutes, and compatible with existing walls of any shape. Finally, a simple demo of an interactive experience has been created to illustrate what can be achieved with the system and the available data structures.
From voice to virtuosity: DDSP-based timbre transfer
(2024) Landarech Mayo, Iker
Timbre is a crucial but elusive feature of music; it originates in the basic physical structure of sound waves but blossoms to touch upon a range of perceptual and social processes that are fundamental to how we derive meaning and emotion from music. This thesis explores timbre transfer with a new Differential Digital Signal Processing (DDSP)-based approach that leverages strong inductive biases without losing the expressive power of deep neural networks and end-to-end learning. By seamlessly integrating classic signal processing synthesizers as differentiable interpretable functions within a neural network, this method achieves high-fidelity generation without relying on large black-box autoregressive models or adversarial losses. This integration results in a domain-appropriate strong inductive bias, enabling a new interpretable representation and modular framework based on parametric signal processors for generative modeling. In this physically grounded approach, fundamental frequency and loudness envelopes are explicitly presented as domain-specific conditioning features to the model, which predicts the timbre via time-varying harmonic amplitudes and filtered noise, yielding an efficient end-to-end backpropagation and reducing model complexity and data requirements. This study presents state-of-the-art results through self-recorded trained models and demonstrates AuraVox, a virtual instrument implementation that enables real-time timbre transfer.
Seguiment de jugadors en el futbol mitjançant tècniques de visió per computador basades en xarxes neuronals
(2024) Corbi Sulé, Carme
El treball de final de grau se centra a investigar el seguiment de jugadors en el futbol a través de tècniques de visió per ordinador. S’explora l’eficàcia de diversos mètodes de seguiment d’objectes utilitzant el conjunt de dades SoccerNet, que ofereix una àmplia gamma d’imatges i vídeos de partits de futbol de diferents escenaris amb anotacions detallades. A través d’una anàlisi exhaustiva, es revisa l’estat actual dels mètodes de seguiment d’objectes, des de mètodes tradicionals fins a les últimes estratègies de xarxes neuronals profundes. S’avaluen els resultats amb mètriques específiques i s’identifiquen els errors i desafiaments trobats. A més, s’obté una millora en els resultats mitjançant l’ús del mètode StrongSORT, centrat a optimitzar la detecció i l’associació d’objectes. Finalment, s’obté un nou model entrenant un detector utilitzant deteccions automàtiques obtingudes a partir de tècniques com Background Subtraction i Human Pose Estimation, amb l’objectiu de millorar la precisió i fiabilitat de les deteccions de jugadors i, per tant, la qualitat global del seguiment de jugadors.
Análisis de redes a partir de una teoría matemática de la consciencia - Teoría de la Información Integrada 4.0
(2024) Cabo Berraco, Daniel
La Teoría de la Información Integrada (IIT, Integrated Information Theory) es una teoría matemática de la consciencia que busca explicar cómo debe configurarse un sustrato físico para emerger como un sistema consciente. A partir de la definición de un conjunto de axiomas que describen la fenomenología de la experiencia, la IIT formaliza unos postulados matemáticamente para establecer las características necesarias con las que un sustrato se convierte en un sistema consciente. Esta teoría puede aplicarse a todo tipo de arquitecturas de red, incluyendo arquitecturas neuronales, computacionales, etc. En el desarrollo de este trabajo, aplicaremos la IIT a redes con diversas características, implementando los postulados propuestos y evaluando cómo afectan a la cantidad y calidad de la experiencia atributos como los grados de conectividad, la distribución de la actividad, la autoexcitación o la autoinhibición. Según esta teoría, la experiencia consciente podrá ser descrita a través de la estructura causa-efecto máximamente irreducible o estructura Φ, especificada por un sustrato en su estado actual, donde la cantidad de experiencia está marcada por su grado de irreductibilidad o información integrada, y la calidad por las distinciones y relaciones que se forman en dicha estructura.
Kobología: virtualización del sintetizador analógico RSF Kobol mediante muestreo
(2022) Castelo Carvajal, Gustavo
El RSF Kobol es un sintetizador análogico monofónico de los años 80 diseñado por la empresa francesa RSF (Ruben y Serge Fernandez), del cual se estima que tras su salida al mercado en 1978 se fabricaron menos de 200 unidades, debido a esto este TFG se propone con la intención de preservar el característico sonido de este sintetizador mediante su virtualización. Este trabajo documenta tanto el análisis previo del sintetizador como el proceso de generación de un banco de muestras grabadas a partir del sintetizador real, así el posterior procesado de esos sonidos y finalmente la construcción del instrumento digital utilizando Kontakt como herramienta principal. El trabajo no solo se centra en la imitación del Kobol original sino también en los componentes artísticos y creativos que tiene la creación de instrumentos musicales digitales. Finalmente se ha creado un instrumento virtual completamente funcional que consigue replicar algunas de las características del Kobol, aunque falla en la imitación del cambio continuo entre formas de ondas de este.
Reconstrucción 3D por partes para objetos dinámicos con regularización temporal
(2022) Pan, Suman
El problema de l’estructura no r ́ıgida del moviment ha despertat un gran inter`es per la comunitat de la visi ́o comunit`aria. Aquest problema ́es recuperar el moviment din`amic 3D d’un conjunt de traject`ories 2D que s’obtenen d’un v ́ıdeo monocular. Malgrat els aven ̧cos realitzats, en els escenaris amb deformacions complexes i un rang de vista restringit, les solucions actuals encara no tenen una soluci ́o robusta. Tanmateix, un enfocament alternatiu en les parts permet proporcionar molts avantatges per resoldre aquestes limitacions respecte al m`etode basat en models globals. En aquest projecte, el m`etode proposat es basa en un algorisme de reconstrucci ́o basat en consens. Com a contribuci ́o, es proposen dues formulacions energ`etiques que explotin un prior de regularitzaci ́o temporal d’un ordre variable i suau, per tal de beneficiar l’estimaci ́o de l’estructura 3D aprofitant les similituds existents entre els marcs consecutius. Els m`etodes proposats s’avaluen amb seq ̈u`encies de moviment hum`a, a partir d’observacions disperses i entrades sorolloses, respectivament. Els resultats mostren que amb aquesta nova restricci ́o de suavitat temporal, es poden obtenir errors de reconstrucci ́o competitius en comparaci ́o amb els m`etodes actuals, alhora que milloren la robustesa de les estimacions.
Desarrollo de una experiencia de realidad aumentada para el nuevo paradigma 'World as Support' mediante el diseño de bloques de código reutilizables. Aplicación sobre la fábrica de Garum de la antigua Barcelona Romana del MUHBA
(2021-07) Pineda Vázquez, Sixto-Sarwed
Con el paso del tiempo, la tecnología ha ido evolucionando en distintas ramas, y una de las que se encuentra en auge hoy en día es la Realidad Aumentada (AR). Desde un punto de vista generalizado dentro del mundo de AR se ha estandarizado el uso del paradigma de Window-on-the-World (WoW) como base para muchos proyectos e investigaciones, pero desde el equipo de Full-Body Interaction Lab (FubIntLab) de la Universidad Pompeu Fabra de Barcelona, se ha planteado un nuevo paradigma, World-as-Support (WaS). Este último plantea la innovación de proyectar toda esa información añadida por AR sobre nuestro mundo físico y no tan solo verla mediante una pantalla, tal y como propone WoW. Con el fin de demostrar el gran potencial que tiene este nuevo paradigma de AR, se empezó el desarrollo del presente proyecto por parte de FubIntLab, donde se pone en práctica una serie de experiencias, basadas en el concepto de WaS, en el Museo de Historia de Barcelona (MUHBA).
Creació d’un mini-joc per a conscienciar als i les adolescents sobre la FoMO (Fear of Missing Out)
(2021-07) González Arnés, Adrià H
Actualment ens trobem en el 2021, un any en el que tothom s’ha hagut de modernitzar, ja que o ens adaptàvem a la situació o ens moríem amb les velles costums. Hem passat d’anar a la universitat dia si i dia també, a fer les classes telemàticament, d’haver d’utilitzar monedes per a comprar la barra de pa, a poder-ho pagar amb el mòbil sense cap problema etc... Doncs, ¿també era el moment d’actualitzar algunes eines educatives, no? La funció d’aquest treball és conscienciar als adolescents de la FoMO (Fear of Missing Out) a través d’un mini-joc d’ordinador i mòbil, i d’aquesta manera canviar la metodologia de les conferencies sobre el perill de les xarxes socials on el jovent no s’implica ni participa. El mini-joc s’ha pogut provar amb adolescents de diferents edats, els quals han participat en una enquesta per a demostrar la seva utilitat. Generalment, els resultats de les enquestes s’ajusten a l’objectiu, ja que acabat el joc els i les adolescents estan més conscienciats i conscienciades de la FoMO.
Eticonsum: aplicación web para la visualización y divulgación de métricas de ética empresarial
(2021-07) Yousfan del Olmo, Ayman
La transformación de nuestro modelo de producción en uno responsable, ético y sostenible es fundamental para reducir la pobreza, la desigualdad y frenar el desastre climático. A pesar de la creciente conciencia entre consumidores y consumidoras, existen varias barreras para que realicen un consumo ético, entre ellas: la falta de información clara, transparente y veraz respecto del impacto y prácticas de las empresas. Este documento recoge el proceso de diseño y desarrollo de una aplicación web que permite visualizar, compartir y explorar información y métricas empresariales con respecto a prácticas éticas, su impacto en el planeta y en las personas. En colaboración con Eticonsum, que nace como una organización sin ánimo de lucro para facilitar esta información en el mercado español, la aplicación permite visualizar sus rankings y todas las demás investigaciones alojadas en Wikirate.org, plataforma de datos abiertos ASG usada por varias ONG, universidades y empresas.
Machine learning-based modelling of expressive style influences among famous cello interpreters
(2021-06) Ovelar Anderson, Sebastian Roberto
The study of the subtle changes musicians introduce when playing a piece has been an area of interest in the past decades. As new and more powerful machines are available, we can build models that imitate these changes. In this project, we have trained models using machine learning methods to predict the expressive performance in terms of timing and dynamics that famous Cellists introduce. Models for timing and dynamics were trained using recordings of famous cello players performing a piece that was segmented into notes. The resulting models were compared to identify expressive performance resemblances among the cello artists by predicting a player’s performance using the expressive performance models of the other ones.
Self-driving RC car using deep learning
(2021-07) Hernández Fuster, Oriol
Self-driving cars have gained popularity in recent years. The increase in computational power coupled with big data has enabled the application of artificial intelligence in such complex areas as driving. Specifically, it is the field of deep learning that has revolutionized these applications. The use of models based on artificial neural networks has achieved great results in a short period of time. In this project we adapt a conventional radio-controlled car into a small-scale autonomous system. The car is able to drive autonomously thanks to the information provided by an RGB monocular camera. We analyze the whole process from assembly, data acquisition, training and test bench.
Design and development of a real-time analytical dashboard for EdTech researchers
(2021-07) Díaz de Burgos Llaberia, Víctor
Educational technology, or EdTech, has grown in importance in the past few years. The adoption of digital tools in class has become widespread in schools, and the analysis of the actions of students is one of the uses. However, the analysis of the activity of professors is still unusual. This final degree project aims to develop an analytical dashboard designed to help EdTech researchers to understand how they prepare their classes. The platform analysed will be edCrumble, it is a web-based platform where professors can design their courses, share them with other users, make comments and take ideas from other designs. To create the dashboard, we will follow an iterative design methodology asking possible end users for their opinion on the tool presented. We will implement a minimum viable product using web development programming languages and Chart.js, a JavaScript library used for data visualization
Analysis of Helmholtz equation from exterior-algebraic multivector Maxwell fields
(2021-07) Diaz Cilleruelo, Iker
Electromagnetism is one of the pillars of our universe. It represents a base for mostly all modern technologies. It is a well defined and understood theory, that has been proven over the years since its inception. Represented by Maxwell’s equations, it describes the behavior and relations of electrically charged particles and electromagnetic fields. While also describing the propagation of electromagnetic waves. However, the equations are expressed by means of vector algebra, and thus are limited to R 1+3. In this thesis, we reformulate the Maxwell’s equations for (k, n, r) time-space dimensions and multivectors of grade r, by means of exterior algebra. We also derive and analyze the Helmholtz equation, alongside the redefinition of the propagation model in those dimensions. With a last point of focus on data transmission via electromagnetic waves.
Decoding natural sounds in early ‘‘visual’’ cortex of congenitally blind individuals
(2021-07) Blanco Borrego, Arnau
In the paper Decoding Natural Sounds in Early “Visual” Cortex of Congenitally Blind Individuals published in 2020 by Lukasz Bola and Petra Vetter, it was observed that it is possible to classify auditory stimuli in the early visual cortex. The aim of this Final Degree Project is to examine and assess the study and modify several steps of the pipeline in order to maximize the classification results.

Navigate

Browse

Recent Submissions