MGM-AE: Self-Supervised Learning on 3D Shape Using Mesh Graph Masked Autoencoders

Area of Science:

• Computer Vision
• Machine Learning
• Geometric Deep Learning

Background:

• Self-supervised learning on 3D mesh data faces challenges with geometric topology and task design.
• Existing methods struggle to effectively model irregular mesh structures.

Purpose of the Study:

• To propose a novel self-supervised learning approach for 3D mesh data.
• To leverage graph masking on mesh faces for effective feature extraction.
• To enhance performance on downstream tasks like shape classification and segmentation.

Main Methods:

• Developed Mesh Graph Masked Autoencoders (MGM-AE) utilizing masked autoencoding.
• Applied graph masking on a mesh graph composed of faces for pre-training.
• Trained and evaluated models under varying masking ratios.

Main Results:

• Achieved state-of-the-art results on shape classification (90.8% accuracy on ModelNet40) and segmentation (78.5 mIoU on ShapeNet).
• Demonstrated superior performance compared to existing mesh encoders.
• Identified optimal masking ratios for best performance.

Conclusions:

• MGM-AE is effective for pre-training on large-scale, unlabeled 3D mesh datasets.
• The proposed method shows significant potential for improving performance on various downstream tasks.
• Graph masking on mesh faces is a viable strategy for self-supervised 3D mesh learning.