HC-PCL: A Hierarchical Cross-Camera Prototypical Contrastive Learning Framework for Unsupervised Object Re-Identification

In recent years, supervised object re-identification (Re-ID) methods have achieved significant progress. However, these methods rely on precise annotations of object instances across different camera views. Due to the substantial appearance variations of the same object captured by different cameras, manual annotation not only consumes considerable human resources but also limits the application of algorithms in real-world scenarios. To address this challenge, this paper proposes an unsupervised training framework based on hierarchical prototype contrastive learning for object re-identification, which eliminates the need for ID labeling of each object instance across multiple cameras. Additionally, we introduce a cross-camera loss to enhance the model's generalization performance in new scenarios. We validate the effectiveness of our method on the VeRi-776 and MSMT17 dataset. Experimental results demonstrate that our approach outperforms state-of-the-art unsupervised object re-identification algorithms in terms of target domain recognition accuracy and cross-domain generalization capability.