CloudMix: Dual Mixup Consistency for Unpaired Point Cloud Completion.

Area of Science:

• Computer Vision
• 3D Data Processing
• Machine Learning

Background:

• Supervised methods struggle with unpaired real-world scans for point cloud completion.
• Existing cross-domain adaptation techniques cause significant information loss from real-world data.
• Need for robust point cloud completion methods that handle domain shift effectively.

Purpose of the Study:

• To develop a point cloud completion method that overcomes limitations of existing cross-domain adaptation techniques.
• To improve the robustness and generalization capability of point cloud completion models on real-world data.
• To preserve fine-grained details and reduce noise in completed point clouds.

Main Methods:

• Proposed a dual mixup-induced consistency regularization strategy integrating source and target domains.
• Implemented mixup in both input and latent feature spaces to enforce consistency in completion predictions.
• Designed a novel density-aware refiner to leverage local context for detail preservation and noise removal.

Main Results:

• The proposed dual mixup method significantly outperforms existing state-of-the-art approaches.
• Demonstrated superior performance on extensive experiments using real-world scans and synthetic unpaired datasets.
• The density-aware refiner effectively preserves fine-grained details and removes outliers from coarse completions.

Conclusions:

• Dual mixup-induced consistency regularization is effective for cross-domain point cloud completion.
• The proposed method offers a robust solution for completing point clouds from unpaired real-world scans.
• Future work can explore further refinements for enhanced detail preservation in 3D reconstruction.