Determining the similarity of deformable shapes

Area of Science:

• Computer Vision
• Cognitive Science
• Pattern Recognition

Background:

• Human and machine shape similarity judgments are crucial for object recognition but remain poorly understood, especially for non-rigid shapes.
• Current methods struggle with shapes not related by rigid transformations and determining articulated part structures.

Purpose of the Study:

• To identify desirable properties of shape similarity methods.
• To evaluate the efficacy of elastic matching in capturing shape similarity, particularly for articulated objects.
• To demonstrate a unified approach for evaluating both smooth and polyhedral shapes.

Main Methods:

• Elastic matching was employed, assessing shape similarity via the sum of local deformations required to transform one shape into another.
• The study focused on comparing local contour properties of shapes.
• Novel results were generated for evaluating smooth and polyhedral shapes using a single method.

Main Results:

• Elastic matching effectively captures similarities in part structure without explicit part decomposition.
• The method demonstrates applicability to both smooth and polyhedral shapes.
• Certain shape similarity effects were identified that are not addressed by current methodologies.

Conclusions:

• Elastic matching provides a robust framework for shape similarity, adeptly handling articulated structures.
• This approach offers a significant advancement for object recognition systems, particularly in complex visual domains.
• Further research is needed to address the limitations of current shape similarity methods identified in this study.