TriCI: Triple Cross-Intra Branch Contrastive Learning for Point Cloud Analysis

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• Self-supervised learning methods for 3D point clouds often use single shared encoders, limiting feature extraction.
• Existing approaches struggle to leverage the full potential of 3D data augmentations.

Purpose of the Study:

• To propose TriCI, a novel triple-branch contrastive learning architecture for enhanced self-supervised feature learning in 3D point clouds.
• To improve the quality and richness of extracted features for downstream tasks.

Main Methods:

• Developed a triple-branch contrastive learning architecture (TriCI).
• Generated three augmented versions of each point cloud sample, creating unique positive pairs.
• Employed distinct encoders for each augmented sample and a novel cross-branch contrastive loss alongside intra-branch loss.

Main Results:

• TriCI demonstrated superior performance in self-supervised learning compared to existing methods.
• Achieved 92.9% accuracy on ModelNet40 for linear SVM evaluation, surpassing competitors by 1.7%.
• Significantly enhanced performance in downstream object classification and part segmentation tasks.

Conclusions:

• The proposed TriCI method effectively addresses limitations of single-encoder structures in contrastive learning.
• TriCI's cross-branch contrastive loss facilitates feature alignment and integration for robust 3D point cloud analysis.
• TriCI shows strong potential for advancing self-supervised learning in 3D computer vision.