Exploiting Robust Memory Features for Unsupervised Reidentification

Jiawei Lian, Da-Han Wang, Xia Du, Yun Wu, Shunzhi Zhu

Abstract: Unsupervised re-identification (ReID) is a task that does not use labels for classification and recognition. It is fundamentally challenging due to the need to retrieve target objects across different perspectives, and coupled with the absence of ID labels for supervision. However, the most severe aspect is that the appearance features of vehicles can be shifted under different viewpoints, which may cause an imbalance in inter-class and intra-class variation. In this work, we propose a fully unsupervised re-identification method applied to the person and vehicle domain. In particular, different from previous methods of cluster pseudo-labelling, we compress the feature dimensions used to generate the labels and improve the quality of the pseudo-label. Then, in the update memory feature module, we introduce the idea of partitioning to construct algorithms for dynamically finding inter-class and intra-class boundaries to improve the robustness of the model. To demonstrate the effectiveness of the proposed method, we conduct experiments on one vehicle dataset (VeRi-776) and one person datasets (MSMT17). Experimental results demonstrate that our method is effective in enhancing the performance of the ReID task, and the proposed method achieves the state-of-the-art performance. The code has been made available at https://github.com/ljwwwiop/unsupervised_reid .