An End-to-End Shape-Preserving Point Completion Network.

Area of Science:

• Computer Graphics
• Computer Vision
• 3-D Reconstruction

Background:

• 3-D point cloud completion is crucial for reconstructing shapes from incomplete data.
• Existing methods often struggle with preserving fine-scale structures and overall shape integrity.

Purpose of the Study:

• To propose an end-to-end shape-preserving point completion network for 3-D point clouds.
• To restore both overall shapes and fine-scale structures of incomplete 3-D objects.

Main Methods:

• An encoder-decoder architecture directly processes incomplete 3-D point clouds.
• A novel encoder captures multi-orientation and multi-scale neighbor information.
• A decoder generates dense and uniform complete point clouds.

Main Results:

• The proposed network effectively restores 3-D shapes, with recovered points closely matching ground truth.
• Achieved state-of-the-art performance, outperforming existing methods in Chamfer distance (CD) and Earth Mover's distance (EMD) errors.
• Demonstrated robustness to noise, varying levels of data incompleteness, and generalization to unseen objects.

Conclusions:

• The end-to-end shape-preserving network offers a robust and effective solution for 3-D point cloud completion.
• The method shows significant improvements in accuracy and generalization capabilities for 3-D shape restoration.