ProtoComp++: Diverse Point Cloud Completion With Controllable Prototype.

Area of Science:

• Computer Vision
• 3D Geometry Processing
• Machine Learning

Background:

• Partial point clouds from sensors pose reconstruction challenges.
• Existing methods struggle with real-world data due to synthetic training gaps.
• Unfamiliar categories and severe incompleteness hinder current point cloud completion models.

Purpose of the Study:

• To develop a robust point cloud completion method with strong generalization.
• To address limitations of synthetic datasets in real-world applications.
• To enhance applicability and performance across diverse point cloud completion tasks.

Main Methods:

• Proposed PrototypeCompletion, a novel prototype-based approach.
• Introduced explicit and implicit prototype integration methods.
• Incorporated language prompts for enhanced training and applicability.

Main Results:

• Demonstrated strong generalization to unseen categories.
• Outperformed state-of-the-art models on PCN and ShapeNet34 benchmarks.
• Excelled in real-world scenarios with various object categories and sensor types.

Conclusions:

• PrototypeCompletion offers improved point cloud completion capabilities.
• The method shows significant potential for real-world 3D reconstruction tasks.
• New benchmark and metric facilitate future research in challenging scenarios.