Deformation Based Curved Shape Representation.

Area of Science:

• Computer Vision
• Computational Geometry
• Machine Learning

Background:

• Analyzing curved shapes is challenging due to variations in scale, location, and rotation.
• Existing methods often struggle with robustness to noise and achieving accurate point correspondence.
• A need exists for a principled representation that handles shape deformations effectively.

Purpose of the Study:

• To introduce a deformation-based representation space for curved shapes.
• To develop a robust shape analysis method using finite dimensional matrix Lie groups.
• To enable accurate shape comparison and clustering through a novel similarity metric.

Main Methods:

• Representing shapes as elements of a finite dimensional matrix Lie group.
• Filtering scale and location variations; identifying rotational variations via equivalence.
• Developing a similarity metric with geodesic solutions and addressing reparametrization invariance.
• Proposing two approaches for point correspondence estimation.
• Adapting k-means clustering for shape analysis within the new representation space.

Main Results:

• The proposed representation demonstrates robustness against local shape perturbations and displacements.
• High precision achieved on Swedish and Flavia leaf datasets.
• Comparable performance to state-of-the-art methods on MPEG-7, Kimia99, and Kimia216 datasets.

Conclusions:

• The matrix Lie group-based representation offers a powerful framework for curved shape analysis.
• The method provides a robust and precise approach for shape recognition and comparison.
• The adaptation of k-means clustering facilitates efficient shape analysis in the proposed space.