Academic Project Proposals

"AI-based perception needs large datasets for training and validation, crucial for applications like autonomous driving and security. Real-world data collection is time-consuming and inconsistent. To address this, digital models (Digital Twins) simulate sensor behavior in controlled environments. At CD6, we're developing a Digital Twin for our multimodal LiDAR devices in Nvidia Omniverse, aiming to enhance data robustness and fidelity.

Extended abstract: Download PDF

Academic Supervisor: Santiago Royo

Supervisor e-mail: santiago.royo@upc.edu

Institution: UOC

co-Supervisor: Josep R. Casas

co-Supervisor e-mail: josep.ramon.casas@upc.edu

Assigned Student Name: Carles Pregonas Graells

Student e-mail: Carles.PregonasGraells@autonoma.cat

pre-Assigned Student Name: Carles Pregonas Graells

Confidential: No

Date: 2024-11-24 14:27:47

"LiDAR devices use light to measure distances but can become blinded by excessive light reflection, causing false detections. This saturation effect, often from retroreflector signals like traffic signs and car plates, currently requires shutting down the sensor in affected areas, which is unacceptable for autonomous vehicles. At CD6, we are developing a custom algorithm to mitigate this risk by studying the radiometric model with retroreflectors and improving software algorithms to prevent this fault.

Extended abstract: Download PDF

Academic Supervisor: Santiago Royo

Supervisor e-mail: santiago.royo@upc.edu

Institution: UB

co-Supervisor: Josep R. Casas

co-Supervisor e-mail: josep.ramon.casas@upc.edu

Assigned Student Name: Xavier Micó Pérez

Student e-mail: Xavier.Mico@autonoma.cat

pre-Assigned Student Name: Xavier Mico Perez

Confidential: No

Date: 2024-11-24 14:28:15

This project focuses on developing deep learning models for automated multi-camera video editing, leveraging a newly introduced unique multimodal dataset curated by professional editors, with the technology of a commercial partner. The goal is to replicate professional editing decisions using advanced audiovisual synchronization and multimodal pattern recognition techniques. Proposal may be remunerated; see extended abstract for details.

Extended abstract: Download PDF

Academic Supervisor: Ismael Benito-Altamirano

Supervisor e-mail: ibenitoal@uoc.edu

Institution: UOC

Confidential: No

Date: 2025-01-16 18:28:47

While significant strides have been made in the application of Multi-modal Large Language Models (MLLMs) to 2D medical images, the exploration of their capabilities in temporal modalities such as fetal ultrasound (FU) video remains underdeveloped. This thesis aims to develop a promptable MLLM for analyzing FU data: extracting, retrieving and measuring relevant information with the help of a virtual assistant.

Academic Supervisor: Dr. Carlos Martin-Isla

Supervisor e-mail: carlos.martinisla@ub.edu

Institution: UB

Confidential: No

Date: 2025-01-27 13:42:44

Data collection for 3D vision is expensive and there are not many realistic high quality datasets publicly available. Researchers are working on 3D simulated data. Being able to generate realistic texture for 3D surfaces, not only provides high quality data but also saves a lot of human working hours. In this project we will explore generative AI models to generate realistic texture for 3D humans.

Academic Supervisor: Sergio Escalera

Supervisor e-mail: sescalera@ub.edu

Institution: UB

co-Supervisor: Meysam Madadi

co-Supervisor e-mail: meysam.madadi@gmail.com

Assigned Student Name: Judith Caldés Nevado

Student e-mail: Judith.Caldes@autonoma.cat

Confidential: No

Date: 2025-01-29 12:07:24

Image style transferring is a domain of research in which the style of a given image is transferred to a target image without changing the main context. An example is virtual cloth try on. Most of SOTA condition their models on a given mask for the inpainting area. Usually the quality of this mask has a direct impact on the results sometimes leading to failures. In this project we will explore and develop ideas for the mask generation in virtual try on.

Academic Supervisor: Sergio Escalera

Supervisor e-mail: sescalera@ub.edu

Institution: UB

co-Supervisor: Meysam Madadi

co-Supervisor e-mail: meysam.madadi@gmail.com

Assigned Student Name: Loc Phat Truong

Student e-mail: LocPhat.Truong@autonoma.cat

Confidential: No

Date: 2025-01-29 12:17:52

Cloth landmark detection is a crucial task in computer vision with applications in fashion, virtual try-on systems, and automated garment analysis. Accurately identifying key landmarks on clothing, such as collars, sleeves, and hems, enables advanced shape modeling, pose estimation, and virtual fitting solutions. This thesis aims to develop an efficient deep learning-based approach to detect and localize clothing landmarks in images.

Academic Supervisor: Sergio Escalera

Supervisor e-mail: sescalera@ub.edu

Institution: UB

co-Supervisor: Meysam Madadi

co-Supervisor e-mail: meysam.madadi@gmail.com

Confidential: No

Date: 2025-01-29 12:26:29

Quantum computing is arising as a novel paradigm that will (re-)define computer vision in a closer future, and big players like Google or IBM provide libraries and python frameworks to faciilitat this transition. CVC's Quantum Machine Learning group (qml.cvc.uab.es) is proposing the analysis of quantum computing approaches for genrerative models. QGANS and QDiffusion algororithms are expected to be explored and compared with their classical conterparts, to dvance in the understanding of the role the QML

Academic Supervisor: Fernando Vilarino

Supervisor e-mail: fernando@cvc.uab.es

Institution: UAB

Confidential: No

Date: 2025-01-29 12:55:24

Knowing the size of fruits (specifically apples) is essential for fruit growers. The aim of this thesis is to develop a new methodology that eliminates the need for reference objects, enabling fruit measurement without external calibration tools. The proposed algorithm will be based on a multi-task neural network. There is an opportunity for the student to receive financial compensation through an employment contract.

Extended abstract: Download PDF

Academic Supervisor: Josep Ramon Morros

Supervisor e-mail: ramon.morros@upc.edu

Institution: UPC

co-Supervisor: Jordi gene Mola

co-Supervisor e-mail: jordi.gene@irta.cat

Confidential: No

Date: 2025-02-03 13:10:55

The goal of this project is to develop a computer vision algorithm that takes a series of images, capturing different but overlapping areas of the retina, and composes them together into a single “wide-field” image. In collaboration with Hospital Sant Joan de Deu, we are interested in developing this tool for images acquired from premature babies, with an eye towards downstream diagnostic models.

Extended abstract: Download PDF

Academic Supervisor: Oscar Camara

Supervisor e-mail: oscar.camara@upf.edu

Institution: UPF

co-Supervisor: Adrian Galdran

co-Supervisor e-mail: adrian.galdran@upf.edu

pre-Assigned Student Name: None

Confidential: No

Date: 2025-02-03 14:46:34

Many disciplines base their studies on multispectral images captured from satellites. The resolution of the provided images and the presence of atmospheric occlusions limits the possible studies of the observed scene. In this project we propose to explore the fusion of multiple satellite images of a same scene, to generate a high-resolution image where the atmospheric distortions have been attenuated

Extended abstract: Download PDF

Academic Supervisor: Daniel Ponsa Mussarra

Supervisor e-mail: daniel@cvc.uab.se

Institution: UAB

Confidential: No

Date: 2025-02-06 17:25:47

Generative models can generate undesired content and they are difficult to control. While fine-tuning and RLHF can fine-tune their behavior, it requires large amounts of data and compute to update model weights. Activation Steering is a cost-effective procedure to control generative models. https://arxiv.org/abs/2410.23054 showed it is possible to alter the style and content of text to image diffusion but its application on image to image diffusion remains unexplored. This will be the topic of the thesis

Academic Supervisor: Pau Rodriguez Lopez

Supervisor e-mail: pau.rodriguez@uab.cat

Institution: UAB

co-Supervisor: Xavier Suau Cuadros

co-Supervisor e-mail: xsuaucuadros@apple.com

Confidential: No

Date: 2025-02-14 12:04:05

In https://arxiv.org/abs/2410.23054, they control language and diffusion models by intervening in all the activations of multiple layers. Here we would explore more targeted (e.g. per token) interventions.

Academic Supervisor: Pau Rodriguez

Supervisor e-mail: pau.rodriguez@uab.cat

Institution: UAB

co-Supervisor: Jordi Gonzalez

co-Supervisor e-mail: jordi.gonzalez@uab.cat

Assigned Student Name: Alex Ferrando De las Morenas

Student e-mail: Alex.FerrandoDeLasMorenas@autonoma.cat

pre-Assigned Student Name: Alex Ferrando

Confidential: No

Date: 2025-02-14 12:09:38

Model merging aims to combine a set of finetuned models into a single model that can perform the task of all the models. This research direction has found much traction in recent years and is also actively investigated in the LAMP group (see e.g. https://arxiv.org/abs/2502.04959). The student will work together with other researchers on this topic. We are especially interested in excellent students which would like to continue with a PhD in this topic. Please contact me for further information.

Academic Supervisor: Joost van de Weijer

Supervisor e-mail: joost@cvc.uab.es

Institution: UAB

Assigned Student Name: Saranne Van Der Linde

Student e-mail: Saranne.VanDerLinde@autonoma.cat

Confidential: No

Date: 2025-02-23 16:18:22

League of Legends is a top e-sports game with global championships. Analyzing players' facial emotions is key to analyze their performance, but their focused expressions appear neutral during the game. This project aims to detect facial micro-expressions to reveal subtle cues, like in poker game. Conducted in collaboration with e-sports start-up Omashu, this study enhances competitive gaming insights.

Academic Supervisor: Jordi Sanchez Riera

Supervisor e-mail: jsanchez@iri.upc.edu

Institution: UPC

co-Supervisor: Cesar Diaz

Confidential: No

Date: 2025-02-24 17:26:32

Several video highlight tools use audio, text, and image analysis but are limited by pre-defined models. To overcome this, we propose leveraging Large Language Models (LLMs) for natural language processing. By analyzing multiple League of Legends tournament recordings, we aim to enhance semantic understanding, intent recognition, and play summarization. This project is in collaboration with e-sports start-up Omashu.

Academic Supervisor: Jordi Sanchez Riera

Supervisor e-mail: jsanchez@iri.upc.edu

Institution: UPC

co-Supervisor: Cesar Diaz

Assigned Student Name: Grigor Grigoryan

Student e-mail: Grigor.Grigoryan@autonoma.cat

Confidential: No

Date: 2025-02-24 17:29:51

Combining data from sensors of multiple vehicles (V) and infrastructure (I) enhances environmental perception, improving safety and traffic via V2V/V2I communication. With limited real-world datasets, synthetic ones are often used for training and testing. This project aims to bridge the gap from virtual to real data by fine tuning with real datasets SOTA systems initially trained with synthetic data.

Academic Supervisor: Josep R. Casas

Supervisor e-mail: josep.ramon.casas@upc.edu

Institution: UPC

Assigned Student Name: Judit Salavedra Pujol

Student e-mail: Judit.SalavedraP@autonoma.cat

pre-Assigned Student Name: Judit Salavedra Pujol

Confidential: No

Date: 2025-02-28 09:43:53

This master thesis proposal is in the context of a project in collaboration with Universidad Complutense de Madrid and UNED. In this project, a dataset of more than 6,000 images depicting the oral cavity has been collected. These images have been categorized into 4 large categories: healthy, benign, potentially malignant and malignant. The goal of this master thesis is to propose new deep learning models and provide an analysis that allows to interpret the models implemented.

Academic Supervisor: Carles Ventura Royo

Supervisor e-mail: cventuraroy@uoc.edu

Institution: UOC

Confidential: No

Date: 2025-02-28 15:41:57

The presence of human beings in the marine environment implies changes in the behavior of the species that inhabit it. To study how this presence influences at different times of the year, 10 cameras will be installed that will record images for 504 hours each. The candidate will study automatic labeling that minimizes incorrect labels (arxiv.org/pdf/2406.14351) and methods that evaluate the contribution to learning of images generated with recent data augmentation techniques (arxiv.org/pdf/2301.02830).

Academic Supervisor: David Masip Rodo

Supervisor e-mail: dmasipr@uoc.edu

Institution: UOC

co-Supervisor: Amalia Manjabacas Soriano

co-Supervisor e-mail: manjabacas@icm.csic.es

Confidential: No

Date: 2025-03-03 11:09:29

This thesis explores Quantum Implicit Representation Networks (QIREN), a quantum-enhanced framework that addresses the limitations of classical Implicit Neural Representations (INRs) in modeling high-frequency details. By integrating quantum circuits with neural networks, QIREN offers improved parameter efficiency and spectral modeling. It outperforms classical methods in signal representation, image super-resolution, and shows promise in 3D tasks like novel view synthesis and scene reconstruction.

Academic Supervisor: Fernando Vilarino

Supervisor e-mail: fernando@cvc.uab.es

Institution: UAB

co-Supervisor: Matias Bilkis

co-Supervisor e-mail: mbilkis@cvc.uab.es

Assigned Student Name: Yeray Cordero Carrasco

Student e-mail: Yeray.Cordero@autonoma.cat

pre-Assigned Student Name: Yeray Cordero

Confidential: No

Date: 2025-03-06 10:34:10

This project aims to develop a multimodal fact-checking dataset and establish baseline verification methods. The dataset comprises unlabeled tweets collected systematically by journalists, focusing on potentially hateful or deceptive content. The study specifically targets misinformation related to immigrants, analyzing both the tweets and their accompanying images. The primary objectives include dataset curation and classification model development for verifying multimodal content.

Extended abstract: Download PDF

Academic Supervisor: Ernest Valveny

Supervisor e-mail: ernest@cvc.uab.es

Institution: UAB

co-Supervisor: Arka Dey

co-Supervisor e-mail: audey@cvc.uab.cat

Confidential: No

Date: 2025-03-06 16:21:36

This project develops AI-driven tools for construction document analysis, including visual similarity models to track blueprint revisions, automatic overlay systems for detecting cross-discipline conflicts, and a “git diff” for technical drawings. It also introduces a specialized vision encoder for construction symbols, an interactive segmentation tool, a diffusion model for architectural layouts, and a knowledge graph linking physical elements to evolving documentation.o

Academic Supervisor: David Vazquez

Supervisor e-mail: david.vazquez@servicenow.com

Institution: UAB

Confidential: No

Date: 2025-03-17 06:04:51

a video sequence. While the U-Net has demonstrated high accuracy in single frames, it prone to present temporal instability in dynamic segmentations. We propose to use Convolutional LSTM cells in U-Net models at 4 different placements: 1) After the U-Net output; 2) In the bottleneck; 3) In each skip connection and 4) in each encoder and decoder blocks. Models will be trained using different combinations of cross-entropy and DICE losses with weights adjusted in a HPO phase.

Extended abstract: Download PDF

Academic Supervisor: Debora Gil

Supervisor e-mail: debora@cvc.uab.es

Institution: UAB

co-Supervisor: carles sanchez

co-Supervisor e-mail: csanchez@cvc.uab.es

Assigned Student Name: María José Millán Meharu

Student e-mail: MariaJose.Millan@autonoma.cat

Confidential: No

Date: 2025-03-17 18:21:19

Look-Up Tables (LUTs) are a key component for image enhancement. LUTs are commonly used to efficiently apply color grading, color enhancement, and other transformations. Recently, different approaches have aimed to replace traditional LUTs with learning-based architectures. However, these approaches are either non-invertible or do not consider context information. Our goal will be to derive an architecture that bridges this gap, being both invertible and considering context information.

Extended abstract: Download PDF

Academic Supervisor: Javier Vazquez Corral

Supervisor e-mail: javier.vazquez@cvc.uab.cat

Institution: UAB

co-Supervisor: Danna Xue

co-Supervisor e-mail: dxue@cvc.uab.cat

Confidential: No

Date: 2025-03-21 13:23:45

This study addresses challenges in H. pylori detection using immunohistochemistry-stained tissue by by leveraging deep learning and uncertainty quantification techniques. Uncertainty will be quantified by analyzing classification inconsistencies across multiple models (CNNs, Vision Transformers) using Bayesian methods, and ensemble learning to distinguish misclassifications caused by artifacts from true ambiguities.

Extended abstract: Download PDF

Academic Supervisor: Debora Gil

Supervisor e-mail: debora@cvc.uab.es

Institution: UAB

co-Supervisor: carles sanchez

co-Supervisor e-mail: csanchez@cvc.uab.es

Assigned Student Name: Pere Lloret Jordi

Student e-mail: Pere.Lloret@autonoma.cat

pre-Assigned Student Name: Pere Lloret

Confidential: No

Date: 2025-03-27 12:59:28

The goal is to achieve a compact representation of dynamic 3D video using streamable 3D Gaussian Splatting (3DGS) .The aim is to enable the streaming of 3DGS scenes over the internet, allowing rendering to occur progressively during data transmission. As more data is transmitted, the visual quality improves, enabling rapid previews of 3DGS over the internet.

Academic Supervisor: Gloria Haro

Supervisor e-mail: gloria.haro@upf.edu

Institution: UPF

co-Supervisor: Minye Wu

co-Supervisor e-mail: minye.wu@kuleuven.be

Assigned Student Name: Merlès Subirà Cribillers

Student e-mail: Merles.Subira@autonoma.cat

pre-Assigned Student Name: Merles Subira

Confidential: No

Date: 2025-03-31 14:59:54

Excessive Central Airway Collapse (ECAC) is a condition where the main airways narrow excessively during exhalation. The diagnosis of ECAC relies on dynamic bronchoscopy, which involves inserting a flexible bronchoscope to visualize the airways in real-time. However, this diagnostic procedure is prone to variability due to subjective visual estimation by clinicians, leading to potential misdiagnosis or inappropriate treatments. To address these limitations...

Extended abstract: Download PDF

Academic Supervisor: Carles Sanchez

Supervisor e-mail: csanchez@cvc.uab.cat

Institution: UAB

co-Supervisor: Debora Gil

co-Supervisor e-mail: debora@cvc.uab.cat

Assigned Student Name: Agustina Ghelfi

Student e-mail: Agustina.Ghelfi@autonoma.cat

Confidential: No

Date: 2025-04-10 16:54:18

In this work we study neural solutions for novel view synthesis of time-varying biological tissues from visual signals in an endoscopy video. Just the video will be exploited without including a priori knowledge to learn a volumetric representation of a non-rigid object by employing sophisticated terms that allow obtaining accurate results.

Academic Supervisor: Antonio Agudo

Supervisor e-mail: aagudo@iri.upc.edu

Institution: UPF

Assigned Student Name: Laura Salort Benejam

Student e-mail: Laura.Salort@autonoma.cat

pre-Assigned Student Name: LAURA SALORT BENEJAM

Confidential: No

Date: 2025-04-11 18:23:04

In multi-modal image registration the goal is to align images corresponding to different modalities. Examples could be an RGB image with a thermal image, or a computer tomography scan with a magnetic resonance image. The goal in this thesis is to extend a recent method for its application in 2D and 3D medical imaging. The method is based on neural networks trained in a self-supervised way.

Extended abstract: Download PDF

Academic Supervisor: Pablo Arias

Supervisor e-mail: pablo.arias@upf.edu

Institution: UPF

co-Supervisor: Gemma Piella Fenoy

co-Supervisor e-mail: gemma.piella@upf.edu

Assigned Student Name: Eudald Elias Vilanova

Student e-mail: Eudald.Elias@autonoma.cat

pre-Assigned Student Name: Eudald Elias

Confidential: No

Date: 2025-04-11 18:50:02

Action anticipation refers to the task of predicting a future action or outcome in a video based on the preceding frames. In our case, we focus on action anticipation in sports videos, specifically basketball, where the goal is to predict which team will secure the rebound after a shot. Additionally, we will explore related tasks such as video classification and action spotting, which serve as simpler baselines for comparison.

Academic Supervisor: Albert Calpes Sintes

Supervisor e-mail: aclapes@ub.edu

Institution: UB

pre-Assigned Student Name: Arnau Barrera Roy

Confidential: No

Date: 2025-04-14 11:23:47

Explainable AI mechanisms might significantly enhance healthcare provider trust by providing transparency into diagnostic recommendations. In this project, we aim to study whether advanced computer vision techniques can reveal how decision support systems can be used to differentiate several types of colonoscopy lesions. This visual transparency could help clinicians to validate AI decisions against their expertise and to increase trustworthiness on the role of AI-based Systems as a valid second opinion.

Academic Supervisor: Jorge Bernal del Nozal

Supervisor e-mail: jorge.bernal@uab.cat

Institution: UAB

pre-Assigned Student Name: Mireia Majo Cornet

Confidential: No

Date: 2025-04-23 10:41:16

Game State Reconstruction aims to recognize the state of a sports game by tracking all players on the field from single-camera video input. Our project focuses on enhancing the GSR-Baseline on the SoccerNet-GSR dataset—an open-source, end-to-end pipeline using state-of-the-art methods in tracking, re-identification, team affiliation, jersey number recognition, pitch localization, and camera calibration—by improving key aspects, such as camera calibration.

Academic Supervisor: Albert Clapes Sintes

Supervisor e-mail: aclapes@ub.edu

Institution: UB

co-Supervisor: Artur Xarles Esparraguera

co-Supervisor e-mail: aclapes@gmail.com

pre-Assigned Student Name: Carme Corbi Sule

Confidential: No

Date: 2025-04-29 12:12:17

This project explores how Document Intelligence (DI) systems can be improved for low-resource and complex scenarios by linking transformer architectures with Ordinary Differential Equations (ODEs). It combines theoretical analysis and practical simulations, applying ODE-based transformers to OCR tasks and analyzing their behavior using dynamical systems theory. This is a co-supervised project between the CRM and the CVC.

Extended abstract: Download PDF

Academic Supervisor: Oriol Ramos Terrades

Supervisor e-mail: oriol.ramos@uab.cat

Institution: UAB

co-Supervisor: David Romero

co-Supervisor e-mail: dromero@crm.cat

Assigned Student Name: Júlia Bujosa Moya

Student e-mail: Julia.Bujosa@autonoma.cat

pre-Assigned Student Name: Julia Burjosa

Confidential: No

Date: 2025-05-05 19:45:02