| Person re-identification(re-ID)aims to find out a certain person from other nonoverlapping camera views accurately when given an image/video of him/her as the query.Previous person re-ID assumes that a person will not change his/her clothes when he/she appears in different cameras and all person data are collected from the same surveillance scenario,thus most re-ID methods will suffer from considerable performance degradation in practical applications.Here we focus on improving the robustness and generality of person re-ID in real-world scenarios from three aspects as follows:1)Cloth-changing person re-ID with clothes template.We propose a new person re-ID task in which a query includes both a person image and a clothes template to find the target person who wears the given clothes in the gallery.A large-scale cloth-changing person dataset which contains 62,328 images of 5,266 identities and 5,266 clothes templates is introduced.Besides,the proposed two-branch network BC-Net is to extract and fuse both biometric feature and clothes feature effectively.Experiments show that the new task has the potential to address the cloth-changing problem.2)Video-based clothchanging person re-ID.This task focus on learning the biometric feature from a video clip of a person.A cloth-changing person dataset named CC-Vi P that contains 101 identities and 5,541 tracklets is also proposed.Tri CE-Net explicitly erases the clothes region in the video clip of a person to learn the biometric feature which can retrieve the target person with different clothes more accurately.3)Domain generalization person re-ID.We propose the Multiple Domain Experts Collaborative Learning(MECL).In MECL,each source domain is associated with a specific domain expert and shares the universal expert.With Domain-Domain and Universal-Domain collaborative learning,domain experts and the universal expert learn mutually.Experiments demonstrate that MECL can improve the model’s generalization capability and outperforms the state-of-the-art methods on most person datasets. |
