Contrastive learning-based joint pre-training for unsupervised domain adaptive person re-identification

Jing Wang, Xiaohong Li, Xuesong Dai, Shuo Zhuang, Meibin Qi

Abstract: The goal of person re-identification (ReID) is to identify and retrieve images of a specific individual across multiple camera views. To reduce manual annotations, the unsupervised domain adaptive (UDA) person ReID task utilizes labeled data from the source domain to learn unlabeled data from the target domain. Existing clustering-based UDA person ReID methods typically involve two steps. First, the initial model is pre-trained using labeled samples from the source domain. Second, the pre-trained model is fine-tuned using pseudo labels generated through feature clustering, along with the corresponding target samples. However, the significant differences between different domains prevent the pre-trained model from extracting valid feature representations of the target samples. Consequently, the generated pseudo labels from clustering are often inaccurate and not conducive to model fine-tuning. We propose a UDA person ReID method based on joint pre-training. Firstly, we jointly pre-train the model using source and target domain data. A contrastive learning method based on the class-level and cluster-level memory dictionary is used to enable the model to learn feature representations with certain generalizations and discrimination, which is beneficial for generating more accurate pseudo labels during the subsequent fine-tuning stage. Secondly, we design a combined offline and online knowledge distillation method for model fine-tuning, which is used to avoid the accumulation of feature bias and further improve the model’s performance. Extensive experiments conducted on three popular datasets demonstrate the effectiveness of our method across multiple benchmarks.