In this thesis, we address two main problems namely the quantitative evaluation of mesh segmentation algorithms and learning mesh segmentation by exploiting the human factor. We propose the following contributions: - A benchmark dedicated to the evaluation of mesh segmentation algorithms. The benchmark includes a human-made ground-truth segmentation corpus and a relevant similarity metric that quantifies the consistency between these ground-truth segmentations and automatic ones produced by a given algorithm on the same models. Additionally, we conduct extensive experiments including subjective ones to respectively demonstrate and validate the relevance of our benchmark. - A new learning mesh segmentation algorithm. A boundary edge function is learned, using multiple geometric criteria, from a set of human segmented training meshes and then used, through a processing pipeline, to segment any input mesh. We show, through a set of experiments using different benchmarks, the performance superiority of our algorithm over the state-of-the-art. We present also an application of our segmentation algorithm for kinematic skeleton extraction of dynamic 3D-meshes.
