Improving the Style Adaptation for Unsupervised Cross-Domain Person Re-identification
Wenyuan Zhang, Li Zhu, Lu Lu · 2020
Most existing person re-identification (Re-ID) methods are based on supervised learning, in which a large amount of labeled data are required for training. However, it remains a challenge task for adapting a model trained in a labeled source domain to an unlabeled target domain, due to the domain gap. To alleviate this problem, we design an unsupervised person style transfer adaptation pipeline for the task of unsupervised domain adaptation (UDA) Re-ID. Following the pipeline, we first apply an image translator to generate style-transferred images. To preserve the ID-related information after translation, we introduce the intra-class similarity and inter-domain diversity, which are crucial properties for Re-ID. In this way, a Cross-domain Similarity Generative Adversarial Network (CSGAN) is proposed to bridge the domain gap. CSGAN is learned by jointly optimizing an image translator and a domain-invariant feature representation network (DIFRN), which constrains the CSGAN to maintain the intra-class similarity and inter-domain diversity during image-image translation. Comparison with current competitive methods demonstrates that the effectiveness of the proposed method under the setting of unsupervised domain adaptation.