CODE: Coherence Based Decision Boundaries for Feature Correspondence.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Feature correspondence is crucial for many computer vision tasks.
• Distinguishing true from false matches is challenging at the local descriptor level, leading to high false positive rates.
• Existing methods often fail with noisy data or wide baselines.

Purpose of the Study:

• To develop a robust method for feature correspondence that overcomes limitations of local descriptor matching.
• To leverage global coherence for improved accuracy and recall in feature matching.
• To enhance feature correspondence systems for challenging scenarios like wide baselines.

Main Methods:

• Proposed a non-linear regression technique to discover a coherence-based separability constraint from noisy matches.
• Embedded this constraint into a correspondence likelihood model.
• Integrated the model into a full feature correspondence system.

Main Results:

• The developed model effectively filters nearest neighbor matches, significantly reducing false positives (over 90% false matches).
• The technique successfully identifies true matches by exploiting global coherence and dominant motion patterns.
• Achieved reliable generation of high-quality correspondences over wide baselines, outperforming previous methods.

Conclusions:

• The proposed coherence-based approach offers a significant advancement in feature correspondence accuracy and robustness.
• This method enables reliable feature matching in challenging conditions where traditional techniques falter.
• The technique has the potential to improve performance in various applications requiring accurate feature matching.