Joint Pre-training Based on Hybrid Contrast and Feature Fusion for Unsupervised Domain Adaptive Person Re-identification

Most existing unsupervised domain adaptive person re-identification methods that rely on pseudo-labels typically follow a two-step process. First, they pre-train the model on the source domain using supervised learning approaches. Afterward, the model is further optimized on the target domain by leveraging the pseudo-labels. However, significant differences between different domains lead to poor model performance when transferring to the target domain. This results in pseudo-labels that are highly noisy, which in turn hinders effective model optimization during the fine-tuning phase. To this end, we propose a pre-training method based on hybrid contrast and feature fusion. Firstly, we incorporate target data into the pre-training phase and guide the model to acquire more robust feature representations through a hybrid contrastive learning approach. Secondly, we design a global-local feature fusion module and utilize the fused features as feature representations for model training, which further boosts the model’s performance. Comprehensive experiments on widely used datasets show that our method consistently outperforms existing approaches.