FAME: 3D Shape Generation Via Functionality-Aware Model Evolution.

Area of Science:

• Computer Vision
• Computational Geometry
• Artificial Intelligence

Background:

• Data augmentation is crucial for improving machine learning models, especially in 3D object recognition.
• Generating diverse and plausible 3D shapes that align with specific functionalities remains a challenge.
• Existing methods often lack the ability to handle cross-functional hybridization or partial shape analysis.

Purpose of the Study:

• To introduce a functionality-aware modeling tool for evolving diverse sets of plausible 3D objects.
• To enable data augmentation, constrained modeling, and open-ended design exploration.
• To address the challenge of evaluating functional plausibility in hybrid or partially defined 3D shapes.

Main Methods:

• Utilizes part recombination to evolve 3D shapes from an initial heterogeneous population.
• Employs a functionality plausibility score, incorporating user preference, for evolutionary selection.
• Develops functionality partial matching to assess plausibility in offspring shapes from cross-category hybridization.

Main Results:

• Generated a variety of plausible hybrid 3D shapes through functionality-aware model evolution.
• Demonstrated that the evolved shapes can complement existing datasets and improve segmentation performance.
• Showcased the tool's capability for constrained modeling and the emergence of novel functional prototypes.

Conclusions:

• The functionality-aware evolution tool effectively generates diverse and plausible 3D objects.
• The approach enhances data augmentation strategies and boosts the performance of 3D data-driven segmentation.
• The tool facilitates both controlled design processes and serendipitous discovery of new object prototypes.