关键词: 换衣行人重识别, 胶囊网络, 矢量胶囊, 标签平滑正则化损失, Circle Loss

Abstract: Abstract: Current research on person re-identification mainly concentrated on short-term situations, on the basis of remaining person’s clothing unchanged. However, in more practical cases, it is common to see that people has higher possibility to change their cloth, which should also be considered within person re-identification models. Hence, a new method of person re-identification of changing clothes based on the joint loss capsule network is conducted. The proposed method is based on the ReIDCaps capsule network model for person re-identification when changing clothes, which uses vector capsules that contain more information than traditional scalar neurons. The length of the vector capsule is used to represent the identity information of the person, and the direction is used to represent the clothing information of the person; Soft Embedding Attention (SEA) is used to prevent model over-fitting; Feature Sparse Representation (FSR) mechanism extracts discriminative features; increases the joint loss of label smoothing regularization cross-entropy loss and Circle Loss to improve the generalization ability and robustness of the model. The method is tested on three datasets Celeb-reID, Celeb-reID-light and NKUP. The experimental results can prove that the method has certain advantages compared with some existing person re-identification methods.

Key words: Keywords: cloth-changing person re-identification, capsules network, vector-neuron capsules, label smoothing cross entropy loss, circle loss