Mesh Convolutional Networks With Face and Vertex Feature Operators.

Area of Science:

• Computer Vision
• Deep Learning
• Geometric Deep Learning

Background:

• Deep learning excels in 1D/2D data but struggles with complex 3D data.
• MeshCNN, a prior method, operates effectively on mesh edges but not other mesh primitives.
• Handling irregular 3D data requires specialized deep learning techniques.

Purpose of the Study:

• To develop novel face-based and vertex-based convolutional operators for 3D mesh deep learning.
• To enhance MeshCNN for broader applicability beyond mesh edges.
• To improve 3D shape analysis and 3D mesh reconstruction tasks.

Main Methods:

• Designed novel face-based convolutional and pooling operators for triangular meshes.
• Developed a new vertex-based convolutional operator for mesh networks.
• Integrated the vertex-based operator into the Point2Mesh model for point cloud reconstruction.

Main Results:

• The face-based architecture achieved state-of-the-art results in mesh classification and segmentation.
• The vertex-based operator significantly reduced training (91%) and inference (20%) times for Point2Mesh.
• No statistically significant performance drop was observed with the vertex-based operator in Point2Mesh.

Conclusions:

• Face-based operators offer superior performance for 3D shape analysis compared to edge-based methods.
• Vertex-based operators provide significant computational efficiency gains in 3D mesh reconstruction.
• These new operators expand the capabilities of deep learning for diverse 3D mesh processing tasks.