NAVE	Networked Augmented Virtual Environment (NAVE) Group

Publication:Danyang Huang, Lei Zhang, Qishuai Diao, Zhong Zhou and Wei Wu. Asymmetric Mutual Learning for Unsupervised Cross-Domain Person Re-identification[C]. Pacific Rim International Conference on Artificial Intelligence. Hanoi, Vietnam, November 8-12, 2021. (CCF rank C) pdf         Unsupervised domain adaptation in person re-identification is a challenging task. The performance of models trained on a specific domain generally degrades significantly on other domains due to the domain gaps. State-of-the-art clustering-based cross-domain methods inevitably introduce noisy labels. The negative effects of noisy labels gradually accumulate during iterative training. Besides, optimizing with conventional triplet loss could make the model stuck in local optima in the late stage of domain adaptation. To mitigate the effects of noisy labels, this paper proposes an asymmetric mutual learning framework which cooperates two models with asymmetric labels. The learned asymmetric information is helpful for the two models to complement with each other. Specifically, we propose a merging clusters algorithm to generate asymmetric labels. We also introduce a similarity weighted loss which can further adapt the model to target domain. Extensive experiments demonstrate that our approach outperforms the state-of-the-art methods on three popular person re-identification datasets. create by admin at 2021-12-13 11:09:55