Frequency Domain Adaptive Modal Alignment for Visible-Infrared Person Re-Identification

Visible–Infrared Person Re-Identification (VI-ReID) is a critical task in intelligent surveillance systems.Existing methods typically rely on complex feature mapping networks to learn shared representations that mitigate modality discrepancies, yet often overlook the intrinsic modality gap between visible and infrared images.Moreover, these methods primarily exploit spatial-domain information and largely underutilize the discriminative potential of the frequency domain. To address these limitations, we propose a novel Frequency Domain Adaptive Modality Alignment (FDAMA) framework, which adopts a hierarchical alignment strategy starting from the frequency domain, performing alignment of representations across different modalities simultaneously at both the image level and the feature level, so as to enforce cross-modal consistency constraints. Specifically, we first introduce the Adaptive Modality Mapping (AMM) module at the image level: this module adaptively filters the amplitude spectra in the frequency domain to effectively reduce the modality gap between visible and infrared images, thereby laying the foundation for subsequent feature-level alignment. Second, we propose an Amplitude Prior-Guided Phase Modulation (APGM) module that incorporates amplitude-derived priors into the modulation of phase features, enhancing their ability to encode semantically consistent cross-modal representations and thereby reducing modality discrepancies at the feature level. Extensive experiments conducted on three widely adopted public benchmarks—SYSU-MM01, RegDB, and LLCM—demonstrate that our proposed FDAMA significantly outperforms existing state-of-the-art (SOTA) methods.