GCFL: Gray-Modality Conversion and Feature Learning for Visible-Infrared Person Re-Identification

Visible-infrared person re-identification (VI-ReID) focuses on retrieving individual images across modalities, specifically from the visible spectrum during daylight and the infrared light in evening. As person images are captured under different spectral conditions, to overcome the significant modality gap is crucial for the VI-ReID method. Existing approaches still suffer from the loss of detailed information, the presence of conversion noise, and the persistent challenge of reducing the modality gap. This research focuses on presenting a novel gray-modality conversion and feature learning framework (GCFL) to tackle the VI-ReID task. Specially, we convert visible and infrared images into grayscale images, and consider them as an auxiliary gray modality by retaining the original modalities. Simultaneously, we introduce conversion embeddings (CEs) to represent conversion noise and conversion-aware weaken loss (CWL) to diminish those noise caused by grayscale transformation. Besides, to minimise the modality gap and to weaken the feature distances in the feature space, gray modality rebalance loss (GMRL) is proposed. Also, gray modality center loss (GMCL) is introduced to cluster the person features of the same identity within the feature space, distinguishing them from features of other identities. Our proposed method demonstrates the competitive performance on the SYSU-MM01 and RegDB datasets, which achieves SOTA VI-ReID performance.