Locality-Guided Global-Preserving Optimization for Robust Feature Matching.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Feature matching is crucial for many computer vision applications.
• Existing methods struggle with outlier rejection and complex transformations.
• Robust identification of correct feature correspondences is essential.

Purpose of the Study:

• To propose a novel framework, LOcality-guided Global-preserving Optimization (LOGO), for effective mismatch removal in feature matching.
• To enhance robustness against outliers and handle complex transformations.
• To improve the accuracy and reliability of feature matching for downstream tasks.

Main Methods:

• A fixed-point progressive approach optimizes a graph-based objective for inlier identification.
• A locality-guided strategy uses local topology consensus to establish an initial high-inlier-ratio set.
• Local affine transformations refine consensus and similarity scores for nodes and edges.

Main Results:

• LOGO demonstrates high effectiveness and robustness in extensive experiments.
• The method achieves competitive performance compared to state-of-the-art techniques.
• LOGO successfully handles complex non-rigid transformations.

Conclusions:

• The proposed LOGO framework significantly improves feature matching accuracy and outlier rejection.
• LOGO shows strong potential for high-level vision tasks like essential matrix estimation, image registration, and loop closure detection.
• The method's geometric approach ensures validity across diverse scenarios.