MESA: Effective Matching Redundancy Reduction by Semantic Area Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Matching redundancy in feature matching leads to inaccurate computations and reduced accuracy.
• Current approaches struggle with fine-grained feature comparison between irrelevant image areas.

Purpose of the Study:

• To reduce matching redundancy and improve feature matching accuracy and efficiency.
• To leverage semantic area matching prior to point matching.

Main Methods:

• Propose MESA (sparse) and DMESA (dense) utilizing Segment Anything Model (SAM) for semantic area identification.
• Develop an Area Graph (AG) for candidate area extraction.
• MESA uses graph energy minimization; DMESA uses dense matching distributions (Gaussian Mixture Model, Expectation Maximization) for efficiency.

Main Results:

• DMESA achieves a nearly five-fold speed improvement over MESA with competitive accuracy.
• Both methods show notable accuracy improvements for nine point matching baselines across diverse datasets.
• Demonstrated generalization and improved robustness against image resolution variations.

Conclusions:

• MESA and DMESA effectively reduce matching redundancy through semantic area matching.
• These methods offer significant improvements in feature matching accuracy and efficiency.
• The proposed approaches show strong potential for real-world applications requiring robust image matching.