Shape L'Âne Rouge: Sliding Wavelets for Indexing and Retrieval

Area of Science:

• Computer Vision
• Computational Geometry
• Data Science

Background:

• Shape representation and retrieval are crucial in fields like medical imaging and molecular biology.
• Existing methods often lack richness, compressibility, or accurate indexing capabilities for complex shape models.

Purpose of the Study:

• To present a novel framework for rich, compressible shape representation.
• To develop an accurate method for indexing and retrieving stored shapes.
• To introduce a natural similarity metric for probability distribution-based shape representations.

Main Methods:

• Representing point-set shapes as the square root of probability densities expanded in a wavelet basis.
• Developing a similarity metric based on the geometry of these probability distributions (arc length on a hypersphere).
• Employing a linear assignment solver for non-rigid alignment of densities via "sliding" wavelet coefficients.

Main Results:

• Demonstrated the framework's utility by successfully matching shapes from the MPEG-7 dataset.
• Provided comparative analysis against other similarity measures, including Euclidean distance shape distributions.
• The proposed representation is both rich and compressible, facilitating efficient shape indexing.

Conclusions:

• The novel wavelet-based framework offers a robust solution for shape representation and retrieval.
• The developed similarity metric accurately captures geometric properties of shapes.
• This approach has significant potential for applications in medical imaging, molecular biology, and remote sensing.