InfoCD: A Contrastive Chamfer Distance Loss for Point Cloud Completion.

Area of Science:

• Computer Vision
• Geometric Deep Learning

Background:

• Point clouds are discrete 3D data points representing geometric surfaces.
• Chamfer distance (CD) is a common metric for point cloud comparison but is sensitive to outliers.
• Existing methods struggle with outlier robustness in point cloud analysis.

Purpose of the Study:

• To propose InfoCD, a novel contrastive Chamfer distance loss.
• To enhance the robustness and efficiency of Chamfer distance for point cloud analysis.
• To improve surface similarity metrics and point cloud completion performance.

Main Methods:

• Developed InfoCD, a contrastive loss that spreads matched points to align point cloud distributions.
• Demonstrated InfoCD's equivalence to maximizing a lower bound of mutual information between geometric surfaces.
• Integrated InfoCD into deep learning frameworks for point cloud completion.

Main Results:

• InfoCD provides a regularized Chamfer distance metric that is robust to outliers.
• Achieved significant improvements in point cloud completion tasks across multiple benchmark datasets.
• Outperformed existing CD-based losses and established new state-of-the-art results.

Conclusions:

• InfoCD offers a computationally efficient and robust alternative to standard Chamfer distance.
• The proposed method enhances the accuracy and reliability of point cloud analysis.
• InfoCD represents a significant advancement in geometric deep learning for 3D data processing.