Modality divergence based evolutionary learning for unsupervised visible-infrared cross-modality person re-identification

Unsupervised visible-infrared cross-modality person re-identification aims to learn cross-modality invariant features between visible and infrared modalities without relying on labeled data. Currently, the state-of-the-art methods optimize cross-modality differences by reducing intra-class gaps while expanding inter-class gaps as the underlying paradigm. However, since the cross-modality intra-class gaps are huge, there must be a large number of inter-class instances between the gaps, and such inter-class instances make cross-modality intra-class instances difficult to get closer to each other in the feature space. To this end, we propose a modality divergence based evolutionary learning framework to optimize the cross-modality intra- and inter-class instance distribution. Specifically, on the one hand, we explore the optimization directions of each cluster in two modalities and make the explored attack and defense clusters perform mutual adversarial evolutionary learning through selection, crossover, and mutation, which produces the optimal inter-class distribution. On the other hand, we explore the intra-class instances with maximum and minimum similarity and perform mutual evolutionary optimization between the maximum and minimum instances, which retains only the modality changes in the intra-class instances to learn cross-modality invariant features. Extensive experiments conducted on datasets for visible-infrared person re-identification demonstrate that the proposed approach outperforms current state-of-the-art methods.