Part-Based Bilinear CNN For Person Re-Identification
Li Li, Jianwu Dang, Yangping Wang, Song Wang, Zhenhai Zhang · 2019
Aiming at the problems of image misalignment and the weak discriminative feature of Person Re-Identification(Re-ID), based on the fine-grained network bilinear CNN, a multipart ReID network is proposed. The branch network is used to learn the part features to reduce the influence of the misalignment problem of the datasets image on the ReID effect, and the compact bilinear pooling(CPB)s used for the fusion of each part of the branch network to generate discriminative feature. Weighted values of block feature and global feature loss are used to optimize the network. The validity of the proposed network structure is verified on the dataset CUHK03 and Market-1501. The results show that the proposed model has higher average recognition accuracy than traditional algorithms and other similar network models.