The triangular mesh has become the most popular representation of three-dimensional (3D) data across various fields, due to its ability to capture complex geometric structures. As the use of this representation grows alongside the increasing interest in leveraging neural networks to automate tasks such as classification, the challenge of enabling neural networks to process the irregular structure of mesh data has become increasingly significant. Convolutional neural networks (CNNs) have achieved remarkable success in classification tasks for two-dimensional (2D) data, sparking interest in extending CNNs to handle mesh data. However, the inherent irregularity of meshes presents a significant challenge for traditional neural network architectures, making it an important area of research to modify CNN architectures to accommodate the unique properties of mesh data.

This thesis investigates the feasibility of employing CNNs for the classification task on triangular meshes. A literature study was conducted to examine existing CNN-based methods for mesh data, identifying successful approaches and addressing the challenges associated with their application. Subsequently, an experimental evaluation was conducted to assess the performance of two established networks for mesh classification: MeshCNN & MeshNet. These networks were evaluated across four datasets, including a custom dataset provided by Configura, for which this thesis was conducted, alongside several widely recognized datasets. The findings of this study demonstrate the potential of mesh CNNs for classification tasks, with MeshNet demonstrating particularly promising results on the custom dataset.