Pose-guided feature fusion for occluded unsupervised person re-identification

Abstract

Unsupervised person re-identification (re-ID) aims to identify the same persons' images across different cameras by training on unlabeled data. In which, how to alleviating the occlusion problem in unsupervised person re-ID is a great challenge. Recently, the work on unsupervised re-ID has achieved substantial progress by clustering on the unlabelled target data or unsupervised domain adaptation. Nevertheless, previous re-ID methods either ignored the occlusion problem or solved it based on extreme assumptions. Therefore, in order to develop a kind of more practical and generalized re-ID methods, this paper propose to alleviate the occlusion problem for the unsupervised model. Firstly, we introduce a poseguided branch to extract the key-points information of person. Then, the global feature extracted by backbone and postural feature obtained by pose-guided branch are fused to fed into the unsupervised system. Finally, the experimental results demonstrate that our identification accuracy has achieved strong performance on the three person re-ID dataset.