Learning Kinematic Structure Correspondences Using Multi-Order Similarities.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Traditional methods match static images using appearance or graph alignment.
• Matching dynamic kinematic structures of heterogeneous objects in videos presents unique challenges.

Purpose of the Study:

• To develop a novel framework for kinematic structure correspondence between articulated objects in videos.
• To enable matching of objects with similar topologies, motions, or both, overcoming limitations of static image matching.

Main Methods:

• The problem is framed as hypergraph matching, incorporating multi-order similarities with normalized weights.
• A structural topology similarity measure is introduced using aggregated topology-constrained subgraph isomorphisms.
• Kinematic correlations and a local motion similarity measure using Riemannian manifold geodesic distance are computed.

Main Results:

• Experiments on synthetic and real data demonstrate the framework's robustness and accuracy.
• The proposed method outperforms various state-of-the-art approaches in kinematic structure correspondence.
• The framework shows versatility across different object types and motion patterns.

Conclusions:

• The novel hypergraph matching framework effectively addresses kinematic structure correspondence in videos.
• This method serves as a foundational component for applications like action recognition and robot motion retargeting.
• The approach offers a significant advancement over existing methods for dynamic object structure matching.