LMR: Learning a Two-Class Classifier for Mismatch Removal.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Feature matching is crucial for vision tasks but current methods rely on transformation models that limit applicability.
• Estimating image transformations for mismatch removal is data-dependent and restricts general use.

Purpose of the Study:

• To develop a general and robust method for mismatch removal in feature matching.
• To cast mismatch removal as a two-class classification problem, learning a universal classifier.

Main Methods:

• Proposed Learning for Mismatch Removal (LMR) as a two-class classification problem.
• Developed a general match representation using consensus of local neighborhood structures via multiple K-nearest neighbors.
• Trained a classifier on a small dataset (10 image pairs, ~8000 matches) for general applicability.

Main Results:

• LMR achieves promising matching results with linearithmic time complexity on arbitrary data.
• Demonstrated generality and robustness across different supervised learning techniques and datasets.
• Outperformed state-of-the-art competitors in feature matching, visual homing, and near-duplicate image retrieval.

Conclusions:

• LMR offers a highly general and robust solution for mismatch removal in feature matching.
• The classification-based approach overcomes limitations of traditional transformation-based methods.
• LMR significantly advances performance in various computer vision applications.