Candidate:
João Ferreira de Carvalho Castro Nunes

Date, Time and Location:
12th December 2025, at 14:00, in Sala de Atos of the Faculdade de Engenharia da Universidade do Porto

President of the Jury:
Pedro Nuno Ferreira da Rosa da Cruz Diniz, Full Professor at the Department of Informatics Engineering of the Faculdade de Engenharia da Universidade do Porto

Members:
Carlos Miguel Fernandes Quental (PhD), Assistant Professor at the Department of Mechanical Engineering, Instituto Superior Técnico, Universidade de Lisboa;
Hugo Pedro Martins Carriço Proença (PhD), Full Professor at the Department of Computer Science, Universidade da Beira Interior;
João Manuel Ribeiro da Silva Tavares (PhD), Full Professor at the Department of Mechanical Engineering, Faculdade de Engenharia, Universidade do Porto (Supervisor);
Luís Paulo Gonçalves dos Reis (PhD), Associate Professor with Habilitation at the Department of Informatics Engineering, Faculdade de Engenharia, Universidade do Porto.

The thesis was co-supervised by Pedro Miguel do Vale Moreira (PhD), Full Professor at the Instituto Politécnico de Viana do Castelo.

Abstract:

Human gait analysis provides critical information on biomechanical function, clinical assessment, and biometric recognition, but achieving accurate and reproducible motion understanding under real-world variability remains a major challenge. Traditional motion capture techniques are dependent on expensive infrastructure and controlled environments, which limit scalability and realworld validity. This thesis addresses these limitations by developing computational methodologies that exploit both RGB and depth information to enable robust, efficient, and fully automatic gait analysis using consumer-grade sensors. The research followed a structured trajectory that encompasses dataset creation, representation design, and methodological innovation. First, an extensive review and comparative analysis of existing vision- and depth-based gait datasets identified gaps in modality diversity, annotation quality, and accessibility. To address these issues, the Gait Recognition Image and Depth Dataset (GRIDDS) was designed, acquired, and publicly released. GRIDDS provides synchronized RGB, depth, silhouette, and 3D skeletal data from 35 participants recorded under controlled conditions, establishing one of the first standardized multi-modal benchmarks for gait analysis and recognition. Building on this foundation, two novel computational gait representations were introduced that fuse two-dimensional appearance cues with three-dimensional skeletal structure to increase robustness to viewpoint, clothing, and carried-object variations. These Gait Skeleton Image (GSI) variants (joint-based and line-based) were integrated within deep learning frameworks and evaluated through extensive experiments, demonstrating competitive and, under certain circumstances, superior performance compared with established appearance-based methods across multiple datasets and covariate conditions. Finally, new methods for gait silhouette interpolation were introduced, combining deterministic geometric reasoning (BRIEF) and bidirectional deep learning (BiSINet) to reconstruct missing frames and enhance temporal coherence. The proposed interpolation techniques significantly improved downstream recognition accuracy and demonstrated strong generalization across datasets and frame-rate conditions. Collectively, the contributions of this work, which span multi-modal data acquisition, robust gait representation learning, and temporal reconstruction, advance the scientific and technological frontiers of human gait analysis, promoting reproducibility, accessibility, and applicability in both clinical and computer vision domains.

Candidate:
Bernardo José Coelho Leite

Date, Time and Location:
17th November 2025, 14:00, Sala de Atos of the Faculdade de Engenharia da Universidade do Porto

President of the Jury:

Pedro Nuno Ferreira da Rosa da Cruz Diniz (PhD), Full Professor in the Department of Informatics Engineering, Faculdade de Engenharia, Universidade do Porto.

Members:

Hugo Ricardo Gonçalo Oliveira (PhD), Associate Professor in the Department of Informatics Engineering, Faculdade de Ciências e Tecnologia, Universidade de Coimbra;

Maria Luísa Torres Ribeiro Marques da Silva Coheur (PhD), Associate Professor in the Department of Informatics Engineering, Instituto Superior Técnico, Universidade de Lisboa;

Luís Paulo Gonçalves dos Reis (PhD), Associate Professor with Habilitation in the Department of Informatics Engineering, Faculdade de Engenharia, Universidade do Porto;

Henrique Daniel de Avelar Lopes Cardoso (PhD), Associate Professor in the Department of Informatics Engineering, Faculdade de Engenharia, Universidade do Porto (Supervisor).

Abstract:

Humans pose questions all the time, and efforts to create AI systems to do the same have been developed. This task, known as Question Generation (QG), is a subfield of natural language generation that aims to automatically produce relevant and grammatically correct questions from a given input, such as text. A key motivation for QG is to support time-consuming tasks like the manual creation of educational questions by teachers. While QG systems have significantly improved, grammatical accuracy alone does not ensure educational value. Consequently, the adoption of QG tools in educational contexts remains limited.

This thesis is driven by three key challenges in QG: (1) the trustworthiness of AI-generated questions; (2) the limited controllability; (3) restricted applicability in less-resourced languages. To address these challenges, we focus on generating open-ended and multiple-choice reading comprehension questions from narrative texts for elementary school students. For challenge 1, we analyze and report the quality of generated questions, identifying both successful and failed cases. For challenge 2, we enhance controllable generation mechanisms by incorporating multiple attributes, such as narrative elements, explicitness, and difficulty, into the generated questions. Challenge 3 is addressed through a special focus on Portuguese, a morphologically rich language that remains underrepresented in QG research.

Our methodology spans from early rule-based and neural approaches to more advanced controllable QG techniques, including fine-tuning, zero- and few-shot prompting with both small and large language models. This offers a comprehensive view of the evolution and performance of QG systems across different stages. We contribute by systematically applying and adapting current QG techniques. We develop case studies that explore controllability and educational relevance, providing comprehensive analyses of question quality, and releasing new QG models and datasets tailored to less-resourced languages such as Portuguese. Evaluation combines automatic metrics with human-centered assessments involving experts, teachers, and students, whose input provides critical insights into the usefulness and effectiveness of the generated questions.

The results show that it is possible to generate well-formulated and answerable questions with controllable attributes. Although machine-generated questions approach the quality of humanauthored ones, semantic issues still arise. In addition, generating MCQs with answer options that are effective for students remains a challenge. These findings highlight the ongoing need for research in educational QG, especially in supporting less-resourced languages and enhancing the reliability of automated generation systems.