A Large-Scale Network Construction and Lightweighting Method for Point Cloud Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Point cloud semantic segmentation is crucial for 3D data understanding.
• Existing methods face challenges with large-scale networks and computational efficiency.
• Terminal deployment requires lightweight and high-performance models.

Purpose of the Study:

• To enhance point cloud semantic segmentation performance.
• To develop an effective lightweighting technique for large-scale networks.
• To address computational cost and terminal deployment demands.

Main Methods:

• Introduced a latent point feature processing (LPFP) module to interconnect base networks (e.g., PointNet++, Point Transformer).
• Proposed a novel point cloud lightweighting method for semantic segmentation networks (PCLN) to compress large-scale networks.
• Utilized selective transfer of structural and attention information for guided network compression.
• Employed feature sampling and aggregation to represent global structure information.

Main Results:

• The LPFP module facilitates feature information transfer and ground truth supervision.
• The PCLN method effectively compresses large-scale networks while retaining performance.
• The proposed method significantly improves performance across various base networks.
• Achieved state-of-the-art results on public and real-world datasets.

Conclusions:

• The novel method significantly enhances point cloud semantic segmentation.
• The lightweighting technique is effective for reducing computational costs and enabling terminal deployment.
• The approach demonstrates broad applicability and superior performance compared to existing methods.