Machine learning for 3D human models remeshing for animation based on a semantic segmentation approach

Polygonal Meshes are the most widespread method of representing 3-dimensional shapes in Computer Graphics, with applications in a plethora of different fields. The need to generate meshes with a clean topology for simulation and animation drives an interest in algorithmic solutions that could automate the work of 3D artists. The goal of the thesis is designing a framework tailor-suited for remeshing realistic human 3D models for animation. To ensure a correct topology a machine learning algorithm is used to extract semantic features from an unstructured triangle human mesh. The semantic features are then converted to feature lines that guide a state-of-the-art remeshing algorithm. The 3D models obtained with the proposed pipeline are compared with the results of the standalone remeshing algorithm to evaluate how the performance improves.