SPC: Self-supervised point cloud completion

Area of Science:

• Computer Vision
• 3D Data Processing
• Machine Learning

Background:

• Point clouds from depth sensors often lack complete shape information.
• Existing completion methods require extensive training data (complete point clouds or multi-view images), limiting real-world applicability.
• High information acquisition costs hinder practical deployment of current point cloud completion techniques.

Purpose of the Study:

• To develop a self-supervised point cloud completion (SPC) method.
• To enable point cloud completion using only single partial point clouds for training.
• To overcome the limitations of existing methods that rely on complete data or multi-view information.

Main Methods:

• An autoencoder-like network architecture with a two-step strategy was developed.
• A compression-reconstruction strategy was employed for learning complete point cloud representations.
• A global enhancement strategy was introduced to prevent overfitting and maintain positional coherence of predicted points.

Main Results:

• The proposed SPC method demonstrated reduced unidirectional Chamfer distance (UCD) and unidirectional Hausdorff distance (UHD) by an average of 2.3 and 2.4 on real-world datasets, respectively.
• The method achieved significant improvements compared to state-of-the-art approaches.
• Application of SPC improved point cloud classification accuracy by an average of 14%.

Conclusions:

• The developed self-supervised point cloud completion method offers a practical solution for reconstructing complete 3D shapes from incomplete data.
• The approach effectively learns from single partial point clouds, reducing reliance on costly data acquisition.
• The method shows high practical value, enhancing both point cloud completion and downstream task performance.