GeoAssemble: A Geometry-Aware Hierarchical Method for Point Cloud-Based Multi-Fragment Assembly.

Area of Science:

• Computer Vision
• Geometric Computing
• 3D Reconstruction

Background:

• Fragment assembly is crucial for cultural relic restoration, medical imaging, and industrial inspection.
• Existing methods struggle with limited feature representation and assembly accuracy.

Purpose of the Study:

• To propose a geometry-aware hierarchical fragment assembly framework (GeoAssemble).
• To improve feature identification, reduce matching ambiguity, and enhance alignment accuracy and stability.

Main Methods:

• Utilized DGCNN for local geometric feature extraction and integrated centroid relative positions for multi-dimensional feature representation.
• Implemented a two-stage matching strategy: global shape similarity coarse matching and local geometric affinity fine matching.
• Developed an auxiliary transformation estimation mechanism using fracture point cloud geometric centers for robust pose parameter initialization.

Main Results:

• The proposed GeoAssemble framework significantly enhances fracture point identification quality.
• The two-stage matching strategy effectively reduces ambiguity in fragment matching.
• The auxiliary transformation estimation improves alignment accuracy and convergence stability.

Conclusions:

• GeoAssemble outperforms baseline methods in matching accuracy and robustness on synthetic and real-world datasets.
• The framework offers a significant advancement in 3D fragment assembly technology.