Unsupervised Person Re-identification via Diversity and Salience Clustering
Xiaoyi Long, Ruimin Hu, Xin Xu · 2021 IEEE International Conference on Systems, Man, and Cybernetics (SMC) · 2021
Obtaining large-scale annotations for person re-identification tasks is very difficult and hard to deploy in real scenarios. Existing unsupervised person re-identification methods focus on one-shot learning and domain adaptation, but they still rely on the labels of the source domain or a small number of sample labels. To solve this problem, we formulate a very simple but effective unsupervised person re-identification method: Diversity and Salience Clustering (DSC). The method generates a stable clustering feature space for unsupervised re-identification by considering the diversity and salience of pedestrian samples. As a pure unsupervised person re-identification model, our approach does not use any camera annotations, pedestrian labels and source domain data. Extensive experiments on two image datasets demonstrate that our proposed method outperforms the state-of-the-art unsupervised person re-identification methods.