CIMatcher: Cross-scale interaction matcher for accurate local feature matching.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Local feature matching is crucial for computer vision tasks like image correspondence.
• Current detector-free methods often neglect multi-scale information, leading to suboptimal feature representations.

Purpose of the Study:

• To propose CIMatcher, a novel detector-free framework to improve local feature matching accuracy.
• To address the limitation of single-scale feature propagation in existing methods.

Main Methods:

• CIMatcher utilizes a multi-scale parallel fusion module (MPFM) to integrate low-level geometric and high-level semantic features.
• A cross-scale feature interaction strategy (CFIS) with an iterative cyclic mechanism promotes feature propagation across scales.

Main Results:

• CIMatcher demonstrates superior performance in homography estimation.
• The framework achieves consistently better results in pose estimation and visual localization tasks.

Conclusions:

• CIMatcher effectively integrates multi-scale information for enhanced local feature matching.
• The proposed framework offers a significant advancement for various computer vision applications requiring accurate feature correspondences.