MeshGraphNet: An Effective 3D Polygon Mesh Recognition With Topology Reconstruction

• Published in: IEEE access Vol. 8; pp. 205181 - 205189
• Main Authors: Song, An Ping, Di, Xin Yi, Xu, Xiao Kang, Song, Zi Heng
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2020; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algebraic topology reconstruction; Computer graphics; Data mining; effectively feature extracting method; Face recognition; Feature extraction; graph convolutional method; Graph neural networks; Image processing; Image reconstruction; Invariance; Medical imaging; mesh recognition; Neural networks; Polygons; Recognition; Shape; Surface reconstruction; Three-dimensional displays; Topology; Uncertainty
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2020.3037236

Three-dimensional polygon mesh recognition has a significant impact on current computer graphics. However, its application to some real-life fields, such as unmanned driving and medical image processing, has been restricted due to the lack of inner-interactivity, shift-invariance, and numerical uncertainty of mesh surfaces. In this paper, an interconnected topological dual graph that extracts adjacent information from each triangular face of a polygon mesh is constructed, in order to address the above issues. On the basis of the algebraic topological graph, we propose a mesh graph neural network, called MeshGraphNet, to effectively extract features from mesh data. In this concept, the graph node-unit and correlation between every two dual graph vertexes are defined, the concept of aggregating features extracted from geodesically adjacent nodes is introduced, and a graph neural network with available and effective blocks is proposed. With these methods, MeshGraphNet performs well in 3D shape representation by avoiding the lack of inner-interactivity, shift-invariance, and the numerical uncertainty problems of mesh data. We conduct extensive 3D shape classification experiments and provide visualizations of the features extracted from the fully connected layers. The results demonstrate that our method performs better than state-of-the-art methods and improves the recognition accuracy by 4-4.5%.