Memory-augmented shuffled meta learning for visible-infrared person re-identification

Visible-infrared person re-identification (VIPR) poses significant challenges due to the inherent differences between visible and infrared images. These differences result in lower similarity among individuals of the same identity across modalities and higher similarity among different identities within the same modality. Existing methods often struggle to effectively address this issue, as they fail to capture global similarity metrics with limited training data, which hinders the model's ability to learn discriminative features. To address these challenges, we introduce a novel approach called memory-augmented shuffled meta (MASM) learning. Our approach is distinguished by two key components: shuffled meta learning (SML) and memory meta learning (MML). SML constructs diverse query and support sets in each training cycle, allowing the model to learn from a wide range of data inputs. Meanwhile, MML leverages historical information stored in memory banks to capture long-term dependencies. This strategic combination of SML and MML not only enhances data utilization but also empowers the model to learn comprehensive global meta metrics, significantly improving its ability to distinguish individuals across modalities. Extensive experiments on the RegDB and SYSU-MM01 datasets validate the effectiveness of our MASM method, demonstrating its superiority over several state-of-the-art approaches.