Collaborative Completion and Segmentation for Partial Point Clouds With Outliers.

Area of Science:

• Computer Vision
• Geometric Deep Learning
• 3D Data Processing

Background:

• Outliers in 3D scanned point clouds degrade performance in geometric tasks.
• Existing methods often require outlier-free data or separate outlier removal steps.
• A synergistic approach between point cloud completion and segmentation is unexplored for outlier handling.

Purpose of the Study:

• To investigate the mutual promotion between point cloud completion and segmentation for robust outlier handling.
• To propose a novel collaborative network, CS-Net, for partial point clouds with outliers.
• To develop a learning paradigm that integrates completion and segmentation without pre-processing.

Main Methods:

• Proposed a collaborative completion and segmentation network (CS-Net) with a cascaded architecture.
• Employed a novel completion network leveraging segmentation labels and farthest point sampling for point cloud purification.
• Utilized KNN-grouping within the completion module for enhanced generation.
• Developed a benchmark dataset for partial point clouds with outliers.

Main Results:

• CS-Net demonstrates significant improvements in outlier robustness compared to existing methods.
• The collaborative approach enhances completion accuracy by utilizing a cleaner, filtered point cloud.
• Segmentation accuracy is improved by leveraging the complete shape inferred by the completion module.
• Extensive experiments validate the effectiveness of the proposed method and dataset.

Conclusions:

• The collaborative mechanism of completion and segmentation networks effectively defeats outliers in 3D point clouds.
• CS-Net offers a robust and accurate solution for processing partial point clouds with inherent outliers.
• The proposed benchmark dataset facilitates further research in robust 3D point cloud analysis.