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

Local feature matching, a fundamental component of plentiful computer vision tasks, aims to establish accurate correspondences between two images. Although current detector-free techniques exhibit impressive performance, they solely rely on single-scale feature propagation while neglecting multi-scale information integration, ultimately yielding suboptimal feature representations for the matching task. To address this limitation, we propose CIMatcher, a new detector-free framework that boosts matching accuracy via cross-scale feature interaction. First, CIMatcher proposes a multi-scale parallel fusion module (MPFM) that adopts a parallel branch structure to effectively integrate low-level geometric features and high-level semantic features, thus providing reliable features for subsequent feature interaction processes. After that, CIMatcher develops a cross-scale feature interaction strategy (CFIS) that utilizes an iterative cyclic mechanism to promote both intra-scale and inter-scale feature propagation, hence extracting discriminative visual descriptors. Extensive experiments indicate that CIMatcher achieves consistently superior performance in all homography estimation, pose estimation, and visual localization tasks.