Unsupervised Cross-domain Person Re-Identification based on Asymmetrical Pyramid Non-local Block

Abstract

The purpose of unsupervised cross-domain (UCD) person re-identification (re-ID) is to adapt the well pre-trained model on the labeled source domain to the unlabeled target domain, which tackles a more realistic problem. However, the network in the existing model cannot fully extract the features of pedestrians, so the results after clustering are not satisfactory. To address this problem, a feature extraction network model with a self-attention mechanism is proposed in this paper in order to improve the feature expression ability. We try to design and optimize the attention mechanism-based feature extraction network and similarity loss function for unsupervised person re-ID to improve the recognition accuracy. On the basis of the baseline network (such as ResNet-50), the self-attention mechanism-asymmetrical pyramid non-local block (APNB) is added to help the network learn richer global feature representation. Besides, the similarity loss function using the Euclidean distance is designed, which shows better performance than the cosine distance. Experimental results show that the proposed method has competitive performance on two public datasets Markket-1501 and DukeMTMC-Re-ID.