CDR-CARNet: Baggage re-identification based on cross-domain robust features and camera-aware re-ranking

To address the challenges of cross-domain distribution inconsistency, large intra-class appearance and viewpoint variations in airport baggage re-identification, this paper proposes CDR-CARNet, which integrates cross-domain robust feature learning, dynamic hard sample mining, and camera-aware re-ranking. Firstly, the integration of Instance-Batch Normalization and Global Context attention mechanisms is employed to alleviate inter-domain shifts. Secondly, Margin Sample Mining Loss is adopted to dynamically select the hardest positive and negative sample pairs, thereby optimizing the decision boundary between samples. Finally, the CA-Jaccard re-ranking strategy is introduced to suppress cross-camera noise interference. Experiments conducted on the MVB dataset demonstrate that CDR-CARNet achieves 87.0% mAP, 86.1% Rank-1, and 84.6% mINP, representing improvements of 4.6%, 4.5%, and 5.9% over the AGW baseline, respectively. The method also significantly outperforms existing mainstream approaches, verifying its practicality and robustness for cross-camera baggage matching in complex airport scenarios.