BGPSeg: Boundary-Guided Primitive Instance Segmentation of Point Clouds.

Area of Science:

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

Background:

• Point cloud primitive instance segmentation is vital for analyzing man-made objects.
• Current methods struggle with accurate instance discrimination near boundaries.
• Over- or under-segmentation issues arise from limitations in feature representations.

Purpose of the Study:

• To propose a novel framework, BGPSeg, for improved point cloud primitive instance segmentation.
• To leverage boundary information for enhanced feature extraction and clustering.
• To address the limitations of existing methods in distinguishing geometric primitives.

Main Methods:

• Introduced a boundary-guided feature extractor incorporating a boundary probability map.
• Utilized boundary-guided sampling and a boundary transformer to improve feature discrimination.
• Developed a boundary-guided primitive clustering module combining boundary clues and geometric features.

Main Results:

• Demonstrated the effectiveness of the BGPSeg framework through extensive experiments.
• Achieved state-of-the-art performance in point cloud primitive instance segmentation.
• Showcased improved segmentation accuracy, especially near geometric boundaries.

Conclusions:

• The proposed BGPSeg framework effectively utilizes boundary information for superior primitive instance segmentation.
• Boundary-guided feature extraction and clustering significantly enhance discrimination between geometric primitives.
• BGPSeg offers a robust solution for understanding geometric shapes in man-made objects from point cloud data.