An Improved Method for Cross-Domain Pedestrian Re-identification

Abstract

Cross-domain pedestrian re-identification (ReID) aims to transfer a model trained on a certain dataset to another unlabeled dataset for testing, which has good robustness. There are two main ways to achieve this so far, one of which is based on generative adversarial network, which aims to let the source domain samples have the style characteristics of the target domain samples, so that reducing the difference in feature distribution between the two domains. The other one is domain adaptation, by mapping the source domain and target domain samples to the same feature space and make the distance in the space as close as possible. This article discusses the reasons for the weak generalization ability of cross-domain models from the macro- and micro-perspectives. Firstly, from the perspective of macro, according to the human parsing model CE2P and the semantic segmentation model DeepLabv3+ , a background segmentation method is proposed. When testing on large differences between two domains such as Market-1501 and MSMT17, the average accuracy of mAP has increased by 1–2%, and the first hit rate top-1 has increased by 3–5%, but it has decreased on small differences between two domains such as Market-1501 and DukeMTMC. Secondly, from the perspective of micro, we combine the advantages of SpCL and cluster contrast, and the labeled samples of the source domain are used to guide the unlabeled samples of the target domain to train. Besides, we select the hardest three samples saved in last epoch when updating, according to the experimental results, compared with the hardest sample in cluster contrast, the MAP and top-1 have better results in three datasets. Finally, we combine with the background segmentation method when testing on the MSMT17, the MAP has increased 0.6%, and top-1 has increased 0.9%. Although the development of unsupervised pedestrian re-recognition is very hot, and the result is close to supervised pedestrian re-recognition, when the model performance reaches to a commanding height, it is more difficult to be improved. However, on the basis of unsupervised, domain adaptation only needs to add the source samples, and it is easy to be improved for more than 2%, so the domain adaptive method still has a larger application prospect.