Cross-Cloud Consistency for Weakly Supervised Point Cloud Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Fully supervised point cloud semantic segmentation demands costly, well-labeled data.
• Existing weakly supervised methods often rely on complex data augmentation or struggle with pseudolabel noise.

Purpose of the Study:

• To develop a cost-effective weakly supervised method for point cloud semantic segmentation.
• To address the challenges of pseudolabel noise and improve learning efficiency.

Main Methods:

• Proposed a cross-cloud consistency method within an expectation-maximization (EM) framework.
• Introduced a pseudolabel selecting (PLS) strategy using cross subcloud consistency in the E-step.
• Implemented cross-scene contrastive regularization in the M-step to reduce noise fitting.

Main Results:

• The method effectively refines pseudolabels and network parameters through alternating learning.
• Cross-cloud constraints enhance learning stability and reduce sensitivity to noisy pseudolabels.
• Achieved significant performance improvements over state-of-the-art weakly supervised methods on three datasets.

Conclusions:

• The proposed cross-cloud consistency method offers a robust and efficient solution for weakly supervised point cloud semantic segmentation.
• The EM framework with PLS and contrastive regularization effectively mitigates pseudolabel noise.
• Demonstrated superior performance and potential for practical applications in 3D data analysis.