Deep mesh labeling via learned semantic boundary guidance. (August 2018)

• Record Type: Journal Article
• Authors: Zhou, Jun; Liu, Xiuping; Cao, Junjie; Wang, Weiming; Yin, Baocai
• Is Part Of: Computer aided design. Volume 101(2018)
• Page Start: 72
• Page End: 81
• Publication Date: 2018-08
• Subjects: Semantic -- Boundary guidance -- Mesh labeling -- CNNs; Computer-aided design -- Periodicals; Engineering design -- Data processing -- Periodicals; Computer graphics -- Periodicals; Conception technique -- Informatique -- Périodiques; Infographie -- Périodiques; Computer graphics; Engineering design -- Data processing; Periodicals; Electronic journals; 620.00420285
• DOI: 10.1016/j.cad.2018.02.001
• Languages: English
• ISSNs: 0010-4485

Abstract

Abstract: We propose a novel method for 3D mesh labeling based on a deep learning approach. We train two deep networks to produce initial labels and semantic boundary maps for test meshes. By using dropout technique, discriminative features can be extracted from our deep networks to improve mesh labeling and boundary detection. Given the detected boundary map, a smoother distance field with closed boundary depiction is calculated for succeeding optimization. Then, based on the initial labels, we obtain the final smooth results through a graph-cut optimization guided by the semantic boundary distance field. With the semantic boundary guidance, labeling is improved distinctly, especially, when large mislabeling regions appear or the boundary of initial labels is not reliable. Furthermore, our algorithm is robust to mesh noise, and can handle mixed dataset with meshes from different categories effectively. Experimental results show that our method outperforms the state-of-the-art methods on public benchmarks. Highlights: We train two deep networks to produce initial labels and semantic boundary maps for 3d meshes. By using dropout technology, distinguishable features can be extracted from our deep networks to improve mesh labeling and boundary detection. Given the detected boundary map, a smoother distance field with closed boundary depiction is calculated for succeeding optimization. Based on the initial labels, we obtain the final smooth results through a graph-cut optimization guided by the semantic boundary distance field. With the semantic boundary guidance, labeling is improved distinctly, especially, when large mislabeling regions appear or the boundary of initial labels is not reliable. Our algorithm is robust to mesh noise, and can handle mixed dataset with meshes from different categories effectively.