Dual-Level Information Transfer for Visible-Thermal Person Re-identification

Visible-thermal person re-identification (VT-ReID) is a challenging pedestrian retrieval problem in the field of security. Due to the intra-modality variations and cross-modality discrepancy caused by different spectrums, it is difficult to extract discriminative features. Existing works are devoted to projecting different-modality features into a shared space, which has weak discriminability and ignores the contextual relationship. In this paper, a novel dual-level information transfer framework is proposed to reduce the modality discrepancy in image level and feature level for VT-ReID. An auxiliary mix-modality is proposed and a mix-visible-thermal (MVT) learning strategy is built to reduce the discrepancy in image level. Firstly, the mix-modality is generated by a mixup scheme which alleviates the direct transfer. Secondly, under the MVT framework, we use ID loss and hetero center triplet loss to guide feature extraction for visible, thermal, and mixed modalities on a one-stream Network. To enhance the robustness of feature extraction, we introduce a graph information transfer module to transfer information across intra-modality and inter-modality in feature level. We build the agent node for modality by using the modality center, where the agent node aggregates the information of all samples in one modality, and then the information from one modality is transmitted to other modalities through the agent nodes. Extensive experimental results on SYSU-MM01 and RegDB datasets show that our method achieves excellent performance.