Dual-Sampling Attention Pooling for Graph Neural Networks on 3D Mesh

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• 3D mesh segmentation is crucial for computer vision and graphics.
• Existing multi-scale deep learning frameworks often neglect vertex receptive field contours, impacting feature scale consistency.
• Current sampling methods have limitations in balancing uniformity and geometric structure preservation.

Purpose of the Study:

• To enhance scale consistency of vertex features in 3D mesh segmentation.
• To improve the preservation of geometric and edge information during mesh sampling.
• To develop a novel attention-based pooling method for effective cross-scale feature transfer in graph neural networks.

Main Methods:

• Utilized uniform sampling to create a multi-scale mesh hierarchy, ensuring feature scale consistency.
• Employed vertex clustering sampling to preserve geometric structure and edge details, mitigating smoothing effects.
• Integrated an attention mechanism to facilitate cross-scale shape feature transfer via a novel graph structure.
• Proposed dual-sampling attention pooling for graph neural networks applied to 3D mesh data.

Main Results:

• The combined sampling methods effectively capture more complete 3D shape information.
• The attention mechanism enables efficient cross-scale feature transfer.
• Experimental results on three datasets demonstrate the high competitiveness of the proposed dual-sampling attention pooling method.
• The approach addresses limitations in previous sampling techniques for 3D mesh analysis.

Conclusions:

• The proposed dual-sampling attention pooling method significantly improves 3D mesh segmentation.
• Combining uniform and vertex clustering sampling enhances feature representation and geometric preservation.
• The attention mechanism is effective for cross-scale feature interaction in graph neural networks for 3D data.
• This work offers a competitive advancement in 3D mesh processing techniques.