Study of remeshing for 3D human models using an AI-based approach

When creating 3D models for real time applications a crucial step in their development is the optimization of the mesh, reducing the number of needed vertices to represent it and in some cases apply a specific topology to it. When dealing with specific kinds of models some special characteristics maybe needed to be applied to said mesh. This task normally involves experts who use multiple hours improving meshes to end with a good final product. Even if some automation tools exist, they are not the norm and normally still require user input to provide the best results. In recent years the use of geometric neural networks has become a topic of interest, and the use of this technology may prove useful automating some of the steps needed to provide good meshes automatically. We focused this thesis in investigating what types of remeshing exist, how they can be applied in such a way that they create a surface with a specific topology and how could we parametrize them in such a way that a machine learning algorithm could tackle them. We divided the whole process into a study on how to remesh, a study on the state-of-the-art on geometrical neural networks and proposed an architecture to provide a frame field that can be used to automate remeshing for human models. The results unfortunately were not precise enough to provide a good final product, but the approach seems promising.