Cupid: Cluster-Based Exploration of Geometry Generators with Parallel Coordinates and Radial Trees

Area of Science:

• Computer Graphics
• Computer Vision
• Data Visualization

Background:

• Geometry generators are crucial for creating 3D assets in games and computer vision.
• Automatic parameter sampling often results in redundant or undesirable geometric shapes.
• Existing methods lack effective tools for exploring the relationship between generator parameters and output shapes.

Purpose of the Study:

• To propose a novel visual exploration approach for geometry generators.
• To enable users to understand parameter sensitivity and identify invalid settings.
• To facilitate the study of abstract parameter spaces and resulting 3D shapes.

Main Methods:

• Developed a composite visualization combining parameter space and 3D shapes.
• Utilized hierarchical clustering to group similar geometric shapes.
• Implemented a parallel coordinates visualization for nested cluster representation.
• Integrated a linked radial tree for cluster hierarchy overview and manipulation.

Main Results:

• Demonstrated effective grouping of similar geometric shapes.
• Enabled iterative drill-down into local shape differences.
• Provided tools for manual cluster splitting and merging.
• Successfully explored the parameter space of a cup generator.

Conclusions:

• The proposed visual exploration approach enhances understanding of geometry generator parameter spaces.
• It aids in identifying and rectifying invalid parameter settings, improving shape generation quality.
• The method offers an intuitive way to navigate complex parameter-output relationships.