Accurate Registration of Cross-Modality Geometry via Consistent Clustering

Area of Science:

• Computer Vision
• Geometric Data Processing
• Machine Learning

Background:

• Unitary-modality geometric data registration is well-established.
• Existing methods struggle with cross-modality registration due to inherent differences between data models.
• A robust solution for aligning dissimilar geometric data is needed.

Purpose of the Study:

• To develop a novel framework for cross-modality geometric data registration.
• To address the limitations of current approaches in handling data from different modalities.
• To improve the accuracy and robustness of point set registration across modalities.

Main Methods:

• Formulating cross-modality registration as a consistent clustering process.
• Employing adaptive fuzzy shape clustering for initial coarse alignment based on structural similarity.
• Optimizing alignment using consistent fuzzy clustering, defining source models as memberships and target models as centroids.
• Investigating the influence of fuzziness in fuzzy clustering and theoretically linking the Iterative Closest Point (ICP) algorithm as a special case.

Main Results:

• The proposed fuzzy clustering method achieves accurate coarse alignment.
• The optimization process significantly enhances robustness against outliers in point set registration.
• Comprehensive experiments on synthetic and real-world data show superior performance compared to state-of-the-art methods.
• The method demonstrates higher accuracy and robustness in cross-modality registration.

Conclusions:

• The consistent fuzzy clustering framework provides an effective solution for cross-modality registration.
• The approach offers improved accuracy and robustness, outperforming existing techniques.
• The study advances the field of geometric data registration by providing a versatile and powerful new method.