An image-computable model of human visual shape similarity.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Computer Vision

Background:

• Human shape perception involves effortless comparison of visual similarities.
• Existing computational models struggle to predict perceived shape similarity.
• A computable model is crucial for understanding neural computations in shape perception.

Purpose of the Study:

• To develop an image-computable model predicting human shape similarity judgments.
• To provide a tool for neuroscientists studying shape perception.
• To investigate the computational basis of human shape representation.

Main Methods:

• Developed 'ShapeComp', a model utilizing over 100 shape features (e.g., area, compactness, Fourier descriptors).
• Trained ShapeComp on a large database (>25,000) of animal silhouettes.
• Validated ShapeComp using novel shapes generated by a Generative Adversarial Network in various perceptual tasks.

Main Results:

• ShapeComp accurately predicts human shape similarity judgments without parameter fitting to human data.
• Incorporating multiple ShapeComp dimensions improves prediction accuracy for small and large sets of shapes.
• ShapeComp surpasses conventional pixel-based metrics and state-of-the-art convolutional neural networks in performance.

Conclusions:

• ShapeComp offers a powerful, image-computable method for predicting human shape similarity.
• The model provides valuable insights into the computations underlying human shape perception.
• ShapeComp can generate perceptually uniform stimulus sets for neuroscience research.