Multiscale Mesh Deformation Component Analysis With Attention-Based Autoencoders.

Area of Science:

• Computer Science
• Geometry Processing
• Shape Analysis

Background:

• Deformation component analysis is crucial for geometry processing and shape understanding.
• Current methods often focus on local, single-scale deformations, which is insufficient for real-world objects with multiscale deformations.

Purpose of the Study:

• To develop an automated method for extracting multiscale deformation components.
• To improve shape understanding and editing by addressing the limitations of single-scale analysis.

Main Methods:

• A novel stacked attention-based autoencoder was proposed.
• The attention mechanism learns to weight multi-scale deformation components in active regions.
• The autoencoder represents deformation components at various scales.

Main Results:

• The proposed method successfully extracts multiscale deformation components automatically.
• Quantitative and qualitative evaluations demonstrate superior performance compared to existing state-of-the-art methods.
• The extracted components enable effective coarse-to-fine shape editing and modeling.

Conclusions:

• The novel stacked attention-based autoencoder effectively addresses the challenge of multiscale deformation analysis.
• This method enhances shape understanding and provides powerful tools for shape editing and creation.