4SNet: Spatial and Spectrum Self-adaptive Synergy Network for Visible-Infrared Person Re-identification

초록

Visible-Infrared Person Re-identification (VI-ReID) is a cross-modality retrieval task that involves searching for a specific pedestrian across different modalities (visible or infrared). Existing methods have explored cross-modality alignment from different perspectives, including spatial-domain cues and frequency-domain modeling. Nevertheless, how to adaptively capture modality-specific frequency characteristics and effectively integrate them into cross-modality feature alignment remains an open challenge. To address this challenge, this work presents a novel and efficient Spatial and Spectrum Self-adaptive Synergy Network (4SNet) for the VI-ReID task, which incorporates an Adaptive Frequency Filter (AFF) module and a Spectrum Synergy Module (SSM) to extract robust cross-modality-invariant features. Specifically, in contrast to traditional fixed convolution-kernel operations, the AFF module reconstructs a set of learnable frequency bases by adopting a dual spatial- and channel attention-guided strategy, which dynamically synthesizes a customized Spectrum filter for each input. Complementarily, the SSM decomposes features into multiple frequency bands, which explicitly decouple identity-shared patterns from modality-specific traits and synergistically enhance them, thereby achieving consistent feature representation across cross-modal pedestrian images at the cross-modality feature alignment level. Extensive experiments have been conducted on the public VI-ReID datasets (SYSU-MM01, LLCM, and RegDB), demonstrating that our proposed method achieves competitive performance when compared to state-of-the-art approaches. Source code implementing 4SNet is publicly available at: https://github.com/dekusaklas-droid/4SNet.

키워드