Supervised, Geometry-Aware Segmentation of 3D Mesh Models

Abstract

Segmentation of 3D model models has applications, e.g., in mesh editing and 3D model retrieval. Unsupervised, automatic segmentation of 3D models can be useful. However, some applications require user-guided, interactive segmentation that captures user intention. This paper presents a supervised, local-geometry aware segmentation algorithm for 3D mesh models. The algorithm segments manifold meshes based on interactive guidance from users. The method casts user-guided mesh segmentation as a semi-supervised learning problem that propagates segmentation labels given to a subset of faces to the unlabeled faces of a 3D model. The proposed algorithm employs Zhou's Manifold Ranking [18] algorithm, which takes both local and global consistency in high-dimensional feature space for the label propagation. Evaluation using a 3D model segmentation benchmark dataset has shown that the method is effective, although achieving interactivity for a large and complex mesh requires some work.