Structural Feature Modulation for Day–Night Cross-Domain Object Re-Identification

Day-night cross-domain object re-identification presents significant challenges due to severe illumination-induced domain gaps. Unlike conventional attention mechanisms that suffer from limited dimensional coverage and rely on single-type normalization strategies, we propose a structural feature modulation (SFM) approach that operates from a modulation perspective. Our SFM approach incorporates a gated batch-layer normalization strategy within the modulation architecture, resulting in the construction of a gated normalization-based modulation (GNM) module. This module effectively leverages the complementary advantages of both batch normalization and layer normalization to balance intra-domain discriminability and cross-domain generalization. Furthermore, we develop a multi-granularity modulation (MGM) module that recalibrates features across both edge-granularity and area-granularity pathways, enabling comprehensive structural modulation. Extensive experiments on the DN-348 and LLCM benchmark datasets demonstrate that our SFM approach consistently outperforms state-of-the-art approaches, achieving notable improvements including a 2.51% mAP increase on the DN-348 dataset.