Title:Homogeneous and Heterogeneous Relational Graph for Visible-infrared Person Re-identification

Abstract:Visible-infrared person re-identification (VI Re-ID) aims to match person images between the visible and infrared modalities. Existing VI Re-ID methods mainly focus on extracting homogeneous structural relationships in an image, i.e. the relations between local features, while ignoring the heterogeneous correlation of local features in different modalities. The heterogeneous structured relationship is crucial to learn effective identity representations and perform cross-modality matching. In this paper, we model the homogenous structural relationship by a modality-specific graph within individual modality and then mine the heterogeneous structural correlation with the modality-specific graph of visible and infrared modality. First, the homogeneous structured graph (HOSG) mines one-vs.-rest relation between an arbitrary node (local feature) and all the rest nodes within a visible or infrared image to learn effective identity representation. Second, to find cross-modality identity-consistent correspondence, the heterogeneous graph alignment module (HGAM) further measures the relational edge strength between local node features of two modalities with routing search way. Third, we propose the cross-modality cross-correlation (CMCC) loss to extract the modality invariance of feature representations of visible and infrared graphs. CMCC computes the mutual information between modalities and expels semantic redundancy. Extensive experiments on SYSU-MM01 and RegDB datasets demonstrate that our method outperforms state-of-the-arts with a gain of 13.73\% and 9.45\% Rank1/mAP. The code is available at this https URL.