Identity Consistency Multi-Viewpoint Generative Aggregation for Person Re-Identification
Yaotao Xiahou, Ning Li, Xiaochao Li · IEEE Transactions on Circuits and Systems for Video Technology · 2023
Person re-identification is the task to retrieve a given person in multiple non-overlapping cameras. Due to viewpoint variation from different cameras, the intra-class variance and inter-class similarity of human images are two critical factors that limit the accuracy of person re-identification. In order to simultaneously resolve these issues, we propose a multi-viewpoint aggregation model, which aims to extend the query method of Re-ID task from single to multi-viewpoints to deal with the various viewpoints and explore inherent gallery information for query optimization. And for constructing multiple auxiliary query images with complementary viewpoints, we design a novel identity consistency pose transfer framework based on a pseudo Siamese structure and trained by a specific Re-ID guided meta-learning pipeline. The goal is to keep the identity consistency between initial query and generated images by enhancing identity-related representation through feature learning and reducing the domain gap between generated and original images. Extensive experiment results indicate our method achieves the Rank-1/mAP performances on Market-1501 (96.62%/93.26%), DukeMTMC-reid (93.45%/87.89%) and CUHK03-labeled (88.26%/87.97%), which outperforms the state-of-the-art based on single viewpoint Re-ID methods.