Quantity-Quality Enhanced Self-Training Network for Weakly Supervised Point Cloud Semantic Segmentation.

Area of Science:

• Computer Vision
• 3D Scene Understanding
• Machine Learning

Background:

• Point cloud semantic segmentation is crucial for interpreting 3D environments.
• Current methods demand extensive annotated data, which is costly and time-consuming to acquire.
• Weakly supervised learning offers a solution by using limited annotations to generate pseudo-labels, but these often lack quantity or quality.

Purpose of the Study:

• To develop an effective weakly supervised method for point cloud semantic segmentation that addresses the limitations of existing pseudo-labeling techniques.
• To improve both the quantity and quality of pseudo-labels for more robust training.
• To achieve competitive performance compared to fully supervised methods.

Main Methods:

• Introduction of the Quantity-Quality Enhanced Self-training Network (Q2E).
• An image-assisted pseudo-label generator to leverage 2D images for extending point cloud pseudo-labels.
• A hierarchical pseudo-label optimizer to refine pseudo-label quality through category grouping.

Main Results:

• Q2E significantly outperforms existing state-of-the-art weakly supervised methods on benchmark datasets (ScanNet-v2, S3DIS, Semantic3D, SemanticKITTI).
• The proposed method achieves performance comparable to fully supervised approaches.
• Q2E secured the top ranking on the ScanNet-v2 benchmark at the time of submission.

Conclusions:

• The Q2E network effectively enhances weakly supervised point cloud semantic segmentation by improving pseudo-label generation and optimization.
• The integration of 2D image information and hierarchical refinement addresses key challenges in pseudo-labeling.
• The method demonstrates strong potential for reducing annotation dependency in 3D scene understanding tasks.