MambaMatch: Establishing Reliable Correspondences via Multi-Scale State Space Model.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Correspondence pruning is crucial for identifying accurate matches between image points, despite outlier disturbances.
• Existing methods like Transformers and graph neural networks face limitations in receptive field size or computational complexity.

Purpose of the Study:

• To introduce MambaMatch, a novel framework for correspondence pruning leveraging state space models.
• To overcome the limitations of existing methods by improving efficiency and accuracy in outlier handling.

Main Methods:

• Proposed MambaMatch, a Mamba-based framework integrating state space models for correspondence pruning.
• Introduced a multi-scale scanning strategy with adaptive clustering for local consensus modeling.
• Developed a Multi-Scale Interaction layer with cross-attention and Gated Feed-Forward Network for feature fusion and discrimination.

Main Results:

• MambaMatch significantly outperforms state-of-the-art methods on multiple benchmarks for two-view geometry estimation.
• Demonstrated robust generalization capabilities across diverse scenarios, tasks, and feature extractors.
• Achieved improved accuracy and efficiency in correspondence pruning compared to existing approaches.

Conclusions:

• MambaMatch represents a pioneering integration of state space models for effective correspondence pruning.
• The proposed multi-scale strategy and interaction layer enhance feature discrimination and local consistency.
• MambaMatch offers a robust and efficient solution for challenging correspondence pruning tasks in computer vision.