Consistent Two-Flow Network for Tele-Registration of Point Clouds.

Area of Science:

• Computer Graphics
• Machine Learning
• Geometric Processing

Background:

• Point cloud registration is crucial for 3D data processing.
• Existing methods fail with small or no overlap between scans.
• Tele-registration, handling arbitrary poses and minimal overlap, remains a challenge.

Purpose of the Study:

• To develop a learning-based technique for robust rigid registration of partial point clouds.
• To enable accurate registration between point clouds with little or no overlap (tele-registration).
• To improve point cloud completion by integrating registration and completion tasks.

Main Methods:

• A novel neural network architecture combining registration and shape completion.
• Simultaneous training using two coupled flows: register-and-complete and complete-and-register.
• Learning a shape prior to complete partial point clouds.

Main Results:

• Achieved robust and reliable tele-registration for point clouds with minimal or no overlap.
• Demonstrated superior performance of individual components in registration and completion compared to state-of-the-art methods.
• Validated the technique on synthetic and real-world partial point clouds.

Conclusions:

• The proposed two-flow training approach significantly enhances tele-registration accuracy.
• The integrated approach leads to better point cloud prediction and completion.
• The method offers a robust solution for challenging point cloud registration scenarios.