Semantic Segmentation to Improve Remeshing of 3D Human Characters

The retopology of a 3D mesh is the process of optimizing the position of the vertices defining its surface to obtain a "cleaner" geometrical representation. Retopology plays a vital role in achieving professional results in various creative and technical fields. Existing automatic methods cannot achieve the results provided by skilled professionals due to the variety of geometrical features characterizing 3D meshes and the large number of conflicting constraints. This research investigates a novel approach that leverages a semantic segmentation of the input model to mimic the human approach of identifying different mesh areas in the retopology process. Edge loops between adjacent semantic regions are used to define the feature lines to be preserved by a field-align remeshing method. The use case consists of the remesh of 3D human-like models to be rigged and animated. The topological quality of the output and its suitability for animation have been evaluated over multiple metrics, and results show that introducing a semantic segmentation step in the retopology pipeline consistently improves the remesh of 3D human models.