Nested Architecture Search for Point Cloud Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Point cloud semantic segmentation (PCSS) is crucial for labeling 3D data points.
• Existing methods struggle to fully utilize context and hierarchical features for representation learning.
• Effective representation learning is vital for accurate PCSS.

Purpose of the Study:

• To develop 'hyperpoint' representations for 3D data points using a nested network architecture.
• To exploit multi-scale, pyramidally hierarchical features for enhanced PCSS.
• To automatically design optimal network architectures for PCSS using a search algorithm.

Main Methods:

• Proposed a nested network architecture to exploit multi-scale, hierarchical features.
• Introduced a PCSS nested architecture search (PCSS-NAS) algorithm.
• Developed the Auto-NestedNet architecture through automated search.

Main Results:

• Auto-NestedNet achieved state-of-the-art performance on four benchmarks (S3DIS, ScanNet, Semantic3D, Paris-Lille-3D).
• The nested architecture effectively addresses the scale-space problem in PCSS.
• Hyperpoint representations capture rich context and hierarchical information.

Conclusions:

• The proposed Auto-NestedNet significantly advances the field of point cloud semantic segmentation.
• Automated architecture search is effective for optimizing PCSS models.
• The 'hyperpoint' representation approach shows great promise for 3D data analysis.