Mesh Neural Networks Based on Dual Graph Pyramids.

Area of Science:

• Computer Vision
• Geometric Deep Learning
• 3D Mesh Processing

Background:

• Deep neural networks (DNNs) are increasingly used for mesh processing but struggle with arbitrary, non-watertight meshes and irregular structures.
• Existing DNNs often require manifold meshes and face challenges in hierarchical feature aggregation.

Purpose of the Study:

• To introduce DGNet, an efficient, effective, and generic deep neural network for processing arbitrary 3D meshes.
• To enable robust mesh analysis and large-scale scene understanding with irregular and defective mesh data.

Main Methods:

• Construction of dual graph pyramids to guide feature propagation across hierarchical levels for downsampling and upsampling.
• A novel convolution method for aggregating local features on hierarchical graphs, utilizing both geodesic and Euclidean neighbors.
• Enabling feature aggregation within local surface patches and across isolated mesh components.

Main Results:

• DGNet demonstrates superior performance on various benchmarks, including ShapeNetCore, HumanBody, ScanNet, and Matterport3D.
• The network successfully handles arbitrary meshes, including those with defects like gaps and non-manifold geometry.
• Experimental results validate DGNet's applicability to both shape analysis and large-scale scene understanding.

Conclusions:

• DGNet provides an efficient and generic solution for deep neural mesh processing, handling complex and arbitrary mesh structures.
• The dual graph pyramid approach and novel convolution effectively address limitations of prior DNNs in mesh analysis.
• DGNet achieves state-of-the-art results, paving the way for more robust 3D data understanding in computer vision.