CAIL: Cross-Modal Vehicle Reidentification in Aerial Images Using the Centroid-Aligned Implicit Learning Network

With the rapid development of autonomous aerial vehicles (AAV) remote sensing equipment, multimodal image data in the remote sensing field have exploded in recent years.In order to effectively alleviate the differences between the three modalities of SAR, visible light, and infrared, we proposed a vehicle reidentification (Re-ID) task based on multimodal aerial images. Compared with traditional Re-ID tasks based on fixed optical cameras, synthetic aperture radar (SAR) has the advantage of being unaffected by lighting and weather conditions, and can provide additional information. However, there are significant geometric distortions and radiometric differences between optical images and SAR images, which limit the effectiveness of multimodal image matching. To address the above issues, we propose a novel centroid-aligned implicit learning (CAIL) network to achieve cross-modal Re-ID. Specifically, we employ a multilevel channel fusion (MTT) module to enhance the adaptability of multimodal encoding channels to dimensional changes, thereby extracting the implicit features from different modalities. Furthermore, by integrating the MTT module into the modality multiple implicit learning (MIL) module, we reduce the modal differences between multimodal images, thus achieving effective alignment between them. Additionally, to optimize CAIL, we propose a modality centroid alignment (MCA) loss to enhance the intraclass feature aggregation capability of multimodal data. MCA dynamically aggregates centroid features by taking modality differences into account, and adopts joint optimization to reduce anomalies in the metric learning process. Our proposed approach achieves significant and satisfactory performance on a cross-modal aerial images dataset, in terms of both mAP and rank-1 accuracy.