Graph convolution-based feature disentanglement for visible-infrared person re-identification

We propose a graph convolution-based disentanglement algorithm that is well-performed in the task of cross-modal person re-identification between visible and infrared images. Given the image of an individual in one modality, the problem to be addressed is whether the same person also appears in images from another modality. To tackle this issue, the main idea of our proposed method is to disentangle image features into modality-related and modality-invariant features, thereby alleviating feature discrepancies across different modal images. Unlike traditional disentanglement methods, our proposed graph convolution-based approach abandons the use of generative adversarial networks and employs attention mechanisms for initial disentanglement, followed by optimization of disentangled features using graph convolution. Comprehensive experimental results on the RegDB dataset and SYSU MM01 dataset demonstrate the superiority of our method in terms of effectiveness and efficiency.