Unsupervised 3D Point Cloud Completion via Multi-View Adversarial Learning.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Machine Learning

Background:

• Scanned 3D point clouds are frequently incomplete due to occlusions.
• Existing self-supervised and weakly-supervised methods struggle with missing data reconstruction.
• Current approaches often require multiple views or overlook intrinsic geometric similarities.

Purpose of the Study:

• To develop a framework for self-supervised point cloud completion using geometric similarities.
• To reconstruct missing regions in incomplete point clouds without 3D ground truth supervision.
• To improve 3D object reconstruction from single-view partial observations.

Main Methods:

• Proposed MAL-UPC framework leveraging region-level and category-specific geometric similarities.
• Introduced a Pattern Retrieval Network to identify and utilize similar geometric patterns.
• Employed multi-view depth map rendering and adversarial learning for geometry refinement.
• Developed a density-aware radius estimation algorithm for anisotropic rendering.

Main Results:

• MAL-UPC effectively completes missing structures in partial point clouds.
• The framework achieves state-of-the-art performance compared to existing self-supervised methods.
• Demonstrated superior results even against some unpaired approaches.
• Successfully reconstructs complete shapes from single-view partial observations.

Conclusions:

• MAL-UPC offers a robust solution for self-supervised point cloud completion.
• Leveraging geometric similarities significantly enhances reconstruction accuracy.
• The proposed method advances the field of 3D data completion from limited observations.