Research On Person Re-identification Algorithm Based On Unsupervised Transfer Learning

Person re-identification(Re-ID)is an image retrieval task aiming to match people across non-overlapping camera views.With the improvement of video surveillance technology,the number of surveillance data has soared,and the labeling work will consume a lot of manpower and material resources.Therefore,the supervised person reidentification methods are limited in practical applications.In order to make the person Re-ID method more scalable,one solution is to formulate the person Re-ID task as an unsupervised domain adaption problem.There are two main challenges in domain adaptation for person Re-ID: Firstly,the target dataset is unlabeled and cannot be fully used.Secondly,since the source and target datasets are often collected from totally different environments,the data distributions of the source and target data are different with a large probability and the domain divergence may cause negative transfer.To address these two issues,existing approaches focus on either creating a cross-data shared representation or estimating the pseudo labels for unlabeled data.However,the former usually ignores the individual characteristics of each domain and the latter always leads to overfitting on such inaccurate pseudo labels.In view of the above two problems,this paper proposes two models to solve the two problems:(1)The self-trained person re-identification algorithm based on re-rankingFirstly,in order to solve the unlabeled problem in the target domain,we adopt a selftraining way.A supervised model is trained on the labeled source domain,and the features of the target domain are extracted using the model.Then we optimize the distance of target samples from the absolute distance between target and target samples and the relative distance between target and source samples,and the more reliable pseudo labels will be generated according to the integrated distance.And the target domain samples with pseudo labels are added to train and update the parameters of model.Then,based on the updated model,the features of the target domain samples are extracted and used to update the pseudo labels to enter the next iteration.The iterations terminate when a stopping criterion is met.Secondly,in order to alleviate the domain divergence,during training,we introduce the domain adversarial loss and one-class classification loss to confuse and close the distribution of the two domains.The focus of this model is to generate more reliable pseudo labels.(2)The person re-identification algorithm based on multi-scale domain adaptive attention learningFirstly,to alleviate domain divergence,a novel domain adaptive attention module(DAAM)is proposed.Given a feature map of any image,the proposed DAAM enforced through the residual mechanism,aims to separate the feature map into the identity-related(IR)feature map and the domain-related(DR)feature map simultaneously.Specifically,the DAAM focuses on capturing the attentive parts of the feature map which is transferable and discriminative,that is,the IR feature map.Simultaneously,the residual parts of the IR are modelled as the DR feature map corresponding to the domain-related information.Then,to obtain more discriminative representation,the proposed DAAM is applied on different feature maps with various scales.Specifically,considering a Res Netlike backbone(such as Res Net-50),a DAAM is introduced to each residual block to extract high-level,mid-level and low-level features respectively.Finally,an IR branch is proposed to the different scale IR feature maps to make sure the IR feature maps discriminative,and the DR branch which follows the DR feature maps employs a crossdataset domain loss.Secondly,to alleviate the influence of unreliable pseudo labels,we consider the pseudo labels as the soft labels,and a novel soft person Re-ID loss is proposed according to the relationship between the training data and the clusters.Specifically,the clusters are regarded as the potential IDs,and the pseudo labels are assigned as possibility distributions rather than definitely IDs.The focus of this model is to alleviate domain divergence.