Unsupervised Dual Transformer Learning for 3-D Textured Surface Segmentation.

Area of Science:

• Computer Vision
• 3-D Geometry Processing
• Machine Learning

Background:

• 3-D texture analysis is crucial for applications like object recognition and material inspection.
• Current methods often rely on supervised learning and global mesh analysis.
• Unsupervised 3-D texture segmentation remains an underexplored area.

Purpose of the Study:

• To propose an original unsupervised framework for segmenting 3-D textures on mesh manifolds.
• To address the challenge of partitioning surfaces into textured and non-textured regions without manual labels.
• To develop a robust method for analyzing local surface variations.

Main Methods:

• A mutual transformer-based system with a label generator (LG) and label cleaner (LC).
• Iterative mutual learning using geometric image representations of mesh facets.
• Binary surface segmentation approach for classifying regions as textured or non-textured.

Main Results:

• The proposed framework achieves effective unsupervised segmentation of 3-D textures.
• Demonstrated superior performance compared to existing unsupervised techniques on diverse datasets.
• Showcased competitive results against supervised methods in segmentation tasks.

Conclusions:

• The developed unsupervised framework offers a significant advancement in 3-D texture analysis.
• The mutual learning approach provides a powerful tool for texture segmentation without annotated data.
• This method holds potential for various applications requiring 3-D surface analysis.