3D Local Feature Learning and Analysis on Point Cloud Parts via Momentum Contrast.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• Self-supervised contrastive learning excels at visual representation learning without labeled data.
• Current 3D local feature learning methods often overlook partial object recognition challenges in real-world scenarios.

Purpose of the Study:

• To develop a momentum contrastive learning framework for discriminative 3D local feature learning from point cloud regions.
• To address the underexplored application of contrastive learning in 3D local feature extraction, particularly for partial observations.

Main Methods:

• Adapted the MoCo (Momentum Contrastive Learning) architecture with PointNet++ as the feature backbone.
• Treated local point cloud parts as fundamental units for contrastive learning.
• Employed augmentation strategies like random dropout and translation for robust feature learning.

Main Results:

• The proposed method effectively learns transferable local features from point cloud regions.
• Demonstrated that approximately 30% of an object's local part is a practical threshold for effective learning under simulated occlusion.
• Achieved comparable downstream classification accuracy with a 16% reduction in training time.

Conclusions:

• The momentum contrastive learning framework is effective for 3D local feature learning from point clouds, especially for partial objects.
• The findings provide insights into practical thresholds for learning from occluded 3D data.
• The method offers an efficient approach to learning robust local features for 3D perception tasks.