Domain adaptive person re-identification via domain alignment and mutual pseudo label refinement
Unsupervised domain adaptive person re-identification refers to transferring knowledge from the labeled dataset to the unlabeled,the need for large amounts of labeled data can be alleviated.The existing methods that address this problem usually use clustering methods to generate pseudo labels.However,those pseudo labels can be unstable and noisy,and can significantly degrade the performance of the methods.In this paper,we propose a novel domain adaptive person re-identification method via domain alignment and mutual pseudo label refinement.Firstly,we extract discriminative feature from the augmented data using a two-branch structure to enrich the feature diversity.Secondly,we design a distributed adversarial domain alignment module to minimize domain differences.Finally,thanks to the complementary relationship between the local and the global features,we establish the consistency between the two kinds of features to refine pseudo labels predicted by the global features,and thus the noise generated by pseudo label clustering is effectively reduced.Extensive experiments demonstrate that the proposed method can achieve remarkable results on popular benchmark datasets for domain adaptive person re-identification.
person re-identificationdomain adaptationdomain alignmentpseudo label refinement
国家科技期刊平台
NSTL
万方数据
