Research On Cross-domain Person Re-identification Based On Pseudo-label Optimization

Person Re-identifification(Re-ID)aims at identifying the target person across different cameras,which has been widely used in many fields such as intelligent surveillance,criminal investigation.Nowadays,many large-scale datasets with manual annotations have promoted the rapid development of this technology.However,even if a model trained well is directly deployed in a new monitoring scenario,the performance of the model will be significantly reduced due to the differences between the domains(Domain Shift).On the one hand,collecting and labeling a large amount of data for all new scenes consumes financial resources.On the other hand,it will bring uncertain label noise.In recent years,many researchers have begun to focus on these cross-domain person re-identification tasks.Among them,the methods based on pseudo labels have obtained leading performance.There are two kind of pseudo labels:the hard pseudo label is an absolute vector with a confidence of 1 in one certain category and a confidence of 0 in other categories,and the soft pseudo label is a smooth vector whose confidence in each category is less than 1.These methods usually first train on the source dataset,then adopt a deep learning mechanism to iteratively generate and optimize pseudo labels to explore the potential identity information of the target domain.The key to the cross-domain person re-identification method based on pseudo label is to mitigate the effects of noisy pseudo labels.Existing leading methods usually use clustering algorithms to generate hard pseudo labels for unlabeled target domains for model training,and the feature extracted by the updated model is re-used for clustering iteratively.This type of method relies heavily on clustering results.Because the number of pedestrian categories in the target domain is unknown,the clustering algorithm itself has shortcomings and the source domain pre-trained model lacks expressiveness in the target domain,the clustering results inevitably contain noise.Moreover,the hard loss used for model learning may continuously amplify noise labels,which will seriously affect model performance.In order to solve this problem,this paper proposes a cross-domain person re-identification method under the guidance of momentum network based on two goals,one is to soften the pseudo label to reduce the influence of noise,and the other is to soften the loss to provide a smooth learning process.The momentum network can be used as a soft supervision.Offline hard pseudo labels and online soft pseudo labels jointly supervise the model,and they can more reasonably represent the identity attributes of pedestrians,and can simultaneously weaken noise label and retain the ability to distinguish pedestrians.The momentum learning network is updated through its past iteration and the backbone network.Taking the past network into consideration can not only prevent the two networks from converging,but also accumulate in time and be more decoupled.This paper indirectly proposes a new way of learning feature distribution:using the traditional triplet loss based on the L2 distance and the newly proposed triplet loss based on the Wasserstein Distance for training,it can narrow the feature distance between the original sample and its positive sample and make the original sample stay far away from its negative sample steadily and smoothly.In the end,the model can effectively reduce the pseudo label noise,thereby improving the performance of pedestrian retrieval.The important task of cross-domain person re-identification method based on pseudo label is to determine whether the generated pseudo label is noisy or not.In the training process,the generated labels are trained as the ground truth label,but there are incorrect predictions among them.If there's a way to determine which of the pseudo labels is noise,the learning efficiency of the model will be improved.Considering that the robust features of adjacent network layers should possess similar characterization capabilities,this paper proposes a cross-domain person re-identification method of selfoptimization based on pseudo-label reliability starting from exploring the robustness of feature representation.First,it uses KL Divergence to quantify the different expressive capabilities of the deep and shallow network features.The result indicates whether the feature expression is strong or not.The reliability of pseudo labels is defined based on this.Therefore,this paper optimizes model learning based on pseudo label reliability by assigning different weights to the loss to suppress the contribution of unreliable labels.A large number of experiments have proved its effectiveness.This paper finally integrates the two methods proposed,and explores the most reasonable way of combining reliability measurement with conventional loss functions.In the end,the integrated network can not only accurately measure the reliability of the pseudo label,but also effectively reduce the noise of the pseudo label.This paper has carried out experimental verification and analysis on the large-scale standard person re-identification datasets such as DukeMTMC-relD,Market-1501 and MSMT17.The experimental results show that the methods proposed in this paper have a good improvement over the existing technology,and verify the effectiveness of these work in the research of unsupervised cross-domain person re-identification.