Credible Dual-X Modality Learning for Visible and Infrared Person Re-Identification

Visible-Infrared person Re-Identification (VI-ReID) is essential for public security. However, it poses a significant challenge due to the distinct reflection frequencies of visible and infrared modalities, leading to a substantial semantic gap between them. A novel modality-transform-based Dual-X method is proposed to narrow the gap between modalities. The modality generators in Dual-X will generate corresponding auxiliary modalities for both visible and infrared modalities, which is achieved through a lightweight channel-level transformation. The newly generated modality images complement the original modal information and are concatenated into the network to facilitate modality-shared and capture modality-specific features. In addition, as softmax is often overconfident on most multi-modal data, an uncertainty estimation algorithm is introduced to quantify the credibility of the model output while providing classification probabilities. By providing reliable uncertainty estimations and reducing uncertainty loss during training, the model’s predictions can be more credible. Extensive experiments were conducted, and the results demonstrated that the proposed approach outperforms state-of-the-art methods by more than 3.7% accuracy on both SYSU-MM01 and RegDB datasets, demonstrating the effectiveness of our approach.