Skeleton-based shape similarity.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Computer Vision

Background:

• Shape similarity is crucial for visual perception, including shape classification and category induction.
• Existing measures of shape similarity lack a universally agreed-upon, principled framework.
• The Bayesian skeleton estimation framework offers a potential basis for a new similarity measure.

Purpose of the Study:

• To derive and validate a novel measure of shape similarity called generative similarity.
• To evaluate generative similarity against existing computational models and human judgments.
• To enhance understanding of how the human visual system perceives and categorizes shapes.

Main Methods:

• Developed generative similarity based on the posterior probability of shapes sharing a common skeletal model.
• Conducted experiments where participants categorized novel 2D and 3D shapes.
• Compared generative similarity predictions against skeletal cross-likelihood, part-based models, and convolutional neural networks.

Main Results:

• Generative similarity demonstrated superior predictive performance in modeling human shape categorization compared to alternative measures.
• The proposed measure effectively captured the nuances of human perception of shape similarity.
• Bayesian skeleton estimation provides a robust foundation for quantifying shape similarity.

Conclusions:

• Generative similarity offers a principled and effective computational approach to understanding human shape perception.
• This research provides insights into the mechanisms underlying shape category induction.
• The findings pave the way for improved models of visual cognition and artificial intelligence systems.