Unsupervised Spectral Mesh Segmentation Driven by Heterogeneous Graphs.

Area of Science:

• Computer Vision
• Geometric Processing
• Computational Geometry

Background:

• Automated mesh segmentation is crucial for 3D data analysis.
• Existing methods often require manual intervention or extensive training data.
• Developing unsupervised techniques remains a significant challenge.

Purpose of the Study:

• To introduce a novel, fully automatic mesh segmentation scheme.
• To leverage heterogeneous graphs for improved segmentation accuracy.
• To provide an unsupervised alternative to current state-of-the-art methods.

Main Methods:

• A spectral framework coupling local geometry and surface patch affinities.
• Construction of a heterogeneous graph combining patch adjacency and dual mesh graphs.
• Partitioning based on individual eigenvector processing of the heterogeneous graph Laplacian, utilizing nodal set and nodal domain theory.

Main Results:

• The proposed unsupervised method achieves superior performance compared to existing unsupervised techniques.
• Segmentation results are comparable to the best supervised approaches on standard datasets.
• Demonstrates the effectiveness of the heterogeneous graph spectral framework.

Conclusions:

• The presented heterogeneous graph-based spectral framework offers a powerful unsupervised mesh segmentation solution.
• This approach advances the field by achieving high accuracy without supervision.
• It provides a viable and efficient alternative for 3D mesh analysis.