Robust Mesh Representation Learning via Efficient Local Structure-Aware Anisotropic Convolution

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• 3D mesh representation learning is crucial for computer vision and graphics.
• Existing graph neural networks for 3D shapes have limitations in representation power due to isotropic filters or predefined local coordinate systems.

Purpose of the Study:

• To propose a novel local structure-aware anisotropic convolutional operation (LSA-Conv) for improved 3D shape representation learning.
• To address the limitations of existing methods in handling irregular 3D shape data.

Main Methods:

• Introduced LSA-Conv, which learns adaptive weighting matrices based on local neighborhood structure.
• Developed LSA-small using matrix factorization to reduce parameter size for high-resolution 3D shapes.
• Incorporated a residual connection with linear transformation to enhance LSA-Conv performance.

Main Results:

• The proposed LSA-Conv and LSA-small models demonstrate significant improvements in 3D shape reconstruction.
• Achieved superior performance compared to state-of-the-art methods in comprehensive experiments.

Conclusions:

• LSA-Conv offers a more powerful approach to 3D shape representation learning by considering local structure.
• The proposed methods effectively overcome limitations of previous graph neural networks for 3D data.