Visible-Infrared Person Re-identification Based on Deep-Shallow Spatial-Frequency Feature Fusion

Visible-Infrared Person Re-identification (VI-ReID) is a challenging task of cross-modality person retrieval. Traditional approaches, hindered by significant inter-modality variations, have predominantly targeted the extraction of shared features in the network's final output layer, often overlooking the valuable shallow-level information. To counter this, the paper introduces a novel framework for VI-ReID, the Deep-Shallow Spatial-Frequency Feature Fusion (DSSF3), which prioritizes the integration of rich, multi-level features. Primarily, the Four-Stream Feature Extraction network (FSFE) is expertly crafted to bridge the gap between visible and infrared images, bolstering the network's fine-grained semantic feature extraction via strategic data augmentation. Concurrently, this paper proposes the Spatial-Frequency Fusion Module (SFFM), which adeptly captures critical spatial and frequency domain details that are commonly neglected during training. The results on two public datasets demonstrate a significant improvement of the proposed method.