Unsupervised person re-identification by hierarchical cluster and domain transfer

Person re-identification (re-ID) has recently been tremendously boosted due to the advancement of deep convolutional neural networks. Unfortunately, the majority of deep re-ID methods focus on supervised, single-domain re-ID task, while less attention is paid on unsupervised domain adaptation. Therefore, these methods always fail to generalize well to real-world scenarios, which have attracted much attention from academia. To address this challenge, we propose a joint unsupervised domain adaptive re-ID method, named HCTL, which is aided by Hierarchical Clustering and Transfer Learning. Specifically, our method performs camera invariance learning using iStarGAN by transferring style of reliable images, which is mined by hierarchical clustering, to the style of other cameras in target domain. During training stage, HCTL integrates TriHard loss on top of ResNet-50 to reduce intra-class variance among dataset and enforce connectedness simultaneously between source domain and target domain. Comprehensive experiments based on Market-1501, DukeMTMC-reID and CUHK03 are conducted, results indicate that our method robustly achieves state-of-the-art performances with only a few reliable samples in target domain and outperform any existing approaches by a large margin.