Learning With Partial and Noisy Correspondence in Graph Matching.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Graph Theory

Background:

• Graph matching methods require precise keypoint correspondences, often unavailable in real-world data.
• Partial correspondence results from occlusions, while noisy correspondence stems from annotation errors (false positives/negatives).

Purpose of the Study:

• To propose the first unified framework addressing both partial and noisy correspondence challenges in graph matching.
• To enhance the robustness and accuracy of graph matching algorithms.

Main Methods:

• A dual-expert cooperative framework integrating Koopmans-Beckmann (KB-QAP) and Lawler's (L-QAP) quadratic assignment programming formulations.
• An align-fuse-refine pipeline utilizing a novel quadratic contrastive loss for alignment and a graph transformer with outlier rejection for fusion.
• Exploiting differential noise resistance of experts to refine correspondences and improve robustness.

Main Results:

• Demonstrated effectiveness on four benchmark datasets against 17 competitive baselines.
• Significant improvements in graph matching accuracy under partial and noisy correspondence scenarios.
• Successful handling of outliers, false positives, and false negatives.

Conclusions:

• The proposed unified framework effectively addresses critical challenges in graph matching.
• The method offers enhanced robustness and accuracy for real-world graph matching applications.
• The approach provides a significant advancement over existing graph matching techniques.