Contrastive Learning for 3D Point Clouds Classification and Shape Completion.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• 3D shape completion and classification are crucial in various applications.
• Current pipelines often rely on AutoEncoders for feature extraction.
• Limited methods effectively capture global features for point cloud datasets.

Purpose of the Study:

• To enhance 3D point cloud shape completion and classification using Self-Supervised Learning.
• To improve global feature learning within AutoEncoder frameworks.
• To demonstrate the generalization capabilities of the proposed method.

Main Methods:

• Integration of contrastive learning with AutoEncoders for global feature extraction.
• Optimization using triplet loss for enhanced feature representation.
• Inclusion of Chamfer distance for local feature learning.
• Evaluation using PointNet classifier on extended class datasets (4 to 10 classes).

Main Results:

• The proposed contrastive AutoEncoder approach improved shape completion and classification accuracy.
• Performance increased from 84.2% to 84.9% for point cloud classification.
• State-of-the-art results were achieved with an expanded 10-class evaluation set.

Conclusions:

• Self-Supervised Learning, particularly contrastive learning within AutoEncoders, significantly boosts 3D point cloud analysis.
• The method demonstrates robust generalization across a larger number of classes.
• This approach offers a promising direction for advanced 3D shape understanding.