Unsupervised person re-identification via K-reciprocal encoding and style transfer

In this paper, we study the unsupervised person re-identification (re-ID) problem, which does not require any annotation information. Our approach considers three aspects in unsupervised re-ID task, i.e., variance across various cameras, label allocation to unlabeled images and hard negative mining. First, an unsupervised style transfer model is adopted to generate style-transferred images with different camera styles, which contributes to reduce the variance across various cameras. Then we apply k-reciprocal encoding method to obtain k-reciprocal nearest neighbors. According to the feature similarity of the probe person with its neighbors, soft pseudo labels are allocated to the probe person iteratively. Due to lack of annotation information to pairwise images, we propose the k-reciprocal nearest neighbors loss (KNNL) to learn discriminative features. Furthermore, a hard negative mining strategy is adopted to improve the accuracy and robustness of our framework. We conduct experiments on three large-scale datasets: Market-1501, DukeMTMC-reID and MSMT17. Results show that our method not only outperforms the state-of-the-art unsupervised re-ID approaches, but also is superior to unsupervised domain adaptation methods (UDA) and semi-supervised learning methods.