Object and background disentanglement for unsupervised cross-domain person re-identification

Abstract

Person re-identification (re-ID) is an important task in many application fields. While most previous works have conducted feature embedding under the supervision of a prior information, it is also true that data collection and annotation in real-world scenarios are very expensive. Moreover, although researchers have borrowed transferring knowledge to alleviate the dependence on labeling, the bias among various domains remains an open problem. Accordingly, motivated by the observation that reducing the style and background differences between domains can promote the generalization capability of the learning model, this paper proposes an unsupervised model, namely, person component decomposition and synthesis generative adversarial network (PCDS-GAN), to minimize the distribution gap among multiple person re-ID datasets. More specifically, we first disentangle the pedestrian image into foreground, background and style features, then use these features to synthesize person images with various backgrounds from the target domain. Finally, the synthesized images are used to train person re-ID models. Comprehensive experiments demonstrate that our model can effectively reduce the domain gap, and also outperforms state-of-the art methods on the Market-1501 and CUHK03 benchmarks.