Domain adaptation with structural knowledge transfer learning for person re-identification

Haojie Liu, Fang Guo, Daoxun Xia

Abstract: Recently, supervised person re-identification (Re-ID) algorithms have achieved great performance on benchmarks. However, it highly depends on labeled training samples and may not generalize well to new domains, limiting the applicability of person Re-ID in practical. To this end, we propose a novel unsupervised domain adaptive approach to transfer the learned knowledge across diverse domains. To address the issue of lacking target domain annotations, we perform a supervised classification task using only labeled source data and share weights of two feature extraction networks. Considering unbalanced data distribution between the source domain and target domain, we then adopt a generative adversarial approach with GAN-based losses to reduce domain discrepancy and further improve Re-ID performance. The entire framework can be trained in an unsupervised manner with standard deep neural networks. Extensive experiments demonstrate that our proposed approach performs favourably against state-of-the-art methods.