Unsupervised Military-Civilian Cross-Domain Vessel Re-Identification Using Improved Momentum Contrast Learning

In this study, we introduce an unsupervised military-civilian cross-domain (UMCCD) benchmark to advance the field of vessel re-identification. By requiring models to train exclusively on unannotated civilian images and test on military images, the UMCCD benchmark rigorously assesses the models' abilities to generalize across significant domain discrepancies. To address these cross-domain challenges, we propose an improved momentum contrast learning (IMCL) method. Our IMCL method enhances the training process by increasing the number of gradient descent optimization layers while reducing the number of momentum update layers. This adjustment improves both feature representation and model robustness, crucially bridging the gap between civilian training data and military test data. Importantly, this research represents the first study to investigate unsupervised military-civilian cross-domain vessel re-identification. Our experimental results show that the IMCL method not only achieves higher accuracy but also demonstrates greater stability when compared to the MoCoV2 momentum contrast method.