Robust Label Rectifying With Consistent Contrastive-Learning for Domain Adaptive Person Re-Identification

Domain adaptive person re-identification (Re-ID) is challenging due to the domain gap between the source and target domains. Existing methods have recently shown great promise by training models with contrastive learning and assigning pseudo labels by clustering, in which a memory bank is utilized to keep features for contrast. However, two main problems lead to sub-optimal generalization ability for existing methods. First, there is no constraint on the updating for memory kept features in existing methods, resulting in inaccurate contrastive learning. Second, the inevitable noisy labels during clustering are usually ignored. To alleviate these problems, we propose a Label Rectifying with Consistent Contrastive-learning (LRCC) framework with two strategies. (1) The consistent contrastive-learning (CC) strategy works with a memory bank which stores the source domain class centroids and all the target domain image features. With the CC strategy, the contrast is conducted across the source and target domains simultaneously. More specifically, we design and maintain consistent clustering during model iteration, thus the classes of memory kept target features are invariable in one epoch. (2) The label rectifying (LR) strategy introduces an auxiliary classifier into the LRCC framework. Thus the pseudo labels are rectified by minimizing the prediction variance between the primary classifier and the auxiliary classifier. To verify the effectiveness of LRCC, we conduct experiments on three public person Re-ID datasets under the domain adaptive setting, DukeMTMC-reID, Market-1501, and MSMT17. The experimental results demonstrate that the proposed LRCC can obtain reliable pseudo labels and achieves state-of-the-art adaptation performance.