Nombre: MATHEUS VIEIRA LESSA RIBEIRO
Fecha de publicación: 05/12/2025
Junta de examinadores:
Nombre |
Papel |
|---|---|
| CLEBESON CANUTO DOS SANTOS | Examinador Externo |
| FILIPE WALL MUTZ | Examinador Interno |
| LUCAS PASCOTTI VALEM | Examinador Externo |
| MARCIA HELENA MOREIRA PAIVA | Examinador Interno |
| RAQUEL FRIZERA VASSALLO | Presidente |
Sumario: This thesis investigates gesture classification using Graph Neural Networks (GNNs) based on the three-dimensional coordinates of skeletal joints. The study addresses challenges related to information propagation in graph structures, particularly the oversquashing phenomenon, which arises from the excessive compression of information during message passing between joints as the number of GNN layers increases. To mitigate these limitations, a graph topology reconfiguration approach based on Ricci curvature and the Augmentations Forman-Ricci Curvature (AFRC) technique was proposed to improve information flow and model performance. The thesis is organized into three proposals, validated on the Chalearn and NTU RGB+D 60 datasets. The first proposal investigated whether information from neighboring joints contributes to the expressiveness of each joint and analyzed the most suitable GNN architecture for such cases. The second proposal addressed the occurrence of the oversquashing effect by introducing an algorithm based on AFRC, which performs random addition and removal of edges in the skeletal topology to enhance the model’s descriptive capacity. Finally, the third proposal employed the StarRGB algorithm as a complementary pipeline applied to joint data. Originally designed for images, this method was adapted to synthesize the temporal evolution of joints across multiple channels, generating a novel gesture representation. Although distinct, the three methodologies share the same general structure, composed of three modules: Graph Constructor, Feature Extractor, and Classifier. The joints are connected according to a given topology and processed by a GNN that extracts a feature vector for each frame. The sequence of these vectors forms a pseudo-image representing the temporal evolution of the gesture, which is then classified by a Convolutional Neural Network (CNN). The experiments demonstrated that the GraphConv architecture achieved the best performance, although increasing the number of layers led to accuracy degradation due to oversquashing. The proposed Ricci
curvature–based algorithm mitigated this effect, enhancing the network’s expressiveness. Furthermore, employing StarRGB as an auxiliary pathway improved classifier performance, establishing it as an alternative for applications involving skeletal joint data.
