Frequency domain adaptive framework for visible-infrared person re-identification

The visible-infrared person re-identification task aims to achieve mutual retrieval between infrared images and visible images. The primary challenge is to learn the mapping of these two modalities into a common latent space. Prior works have mainly focused on network feature extraction, but have overlooked the local information of high-frequency channel features, the global information of low-frequency channel features, and the interaction effects between them, all of which are crucial for effectively aligning feature spaces and enhancing cross-modal recognition accuracy, robustness, and overall performance. To address this issue, we propose a frequency domain adaptive framework. Specifically, we designed the frequency domain adaptive encoder to achieve frequency domain adaptation. And the diverse wise embedding was designed to efficiently extract multi-scale features with fewer parameters. Additionally, we proposed the similarity distance clustering strategy, which reduces the large gaps between different modalities by minimizing the KL divergence between visible-infrared similarity distributions images and the normalized label clustering distributions. Our method has been proven superior on two public datasets and achieves state-of-the-art performance on the RegDB dataset.