Research On Metric Learning And Evaluation Metric For Person Re-identification

Person re-identification technology can effectively facilitate investigators to find and track suspects in massive amounts of video,thereby improving the crime-solving rate of public security departments,which has great research significance and practical application value.In recent years,state-of-the-art person re-identification methods have achieved excellent accuracy on public datasets,even beyond human discriminative capability.However,there are still some challenges in existing person re-identification research,which are as follows: limited metric space of optimization strategy,limited design efficiency of metric learning method,and limited evaluation aspect of evaluation metric.These three aspects of the research challenges are highlighted below:(1)During the training process of existing person re-identification algorithms,a piecewise optimization strategy of partial samples within a batch is used,leading to a local optimization of samples in the metric space.(2)Existing metrics are designed with the human-designed mode,which requires a lot of expertise accumulation and resource consumption to design,resulting in low design efficiency.(3)Existing evaluation metrics are unable to perform a comprehensive and integrated evaluation of retrieval and verification when measuring person re-recognition algorithms,resulting in an imperfect evaluation of the algorithms.To this end,this paper conducts research on metric learning and evaluation metric methods for person re-identification,focusing on three aspects: optimization strategy of metric learning methods,automated design of metric learning methods,and the design of multi-scenarios evaluation metrics.The innovative results have been achieved as follows:(1)Batch Sample Globalization-oriented Metric Learning Method for Person Re-identificationTo address the problem that existing pedestrian re-identification metric learning methods have limited metric space of optimization strategy,consistent metric optimization between model training and evaluation is considered.However,due to the non-differentiable nature of evaluation metrics,it is difficult to train the metric learning methods.This paper investigates person re-identification metric learning methods with batch sample globalization,and achieves the design of metric optimization methods based on evaluation metrics,which can then guide the model training properly.Experimental results show that the proposed method not only optimizes the inter-class distance distribution,but also maintains the intra-class similarity structure.It was validated on three public person re-identification datasets,i.e.,Market1501,CUHK03,and MSMT17,and outperformed state-of-the-art metric learning methods.(2)Automated Designed Metric Learning Method for Person Re-identificationTo address the problem that the design efficiency of existing person re-identification metric learning methods is limited,the automated design of metric learning methods is explored.The automated search of metric learning methods is achieved with the help of automatic machine learning,and the automated design of metric learning methods is completed.However,the metric learning methods derived from the automatic search have no fixed form and poor scalability,depending heavily on the network structure and datasets,which makes the final metric learning methods impracticable.In this paper,we propose an automated designed metric learning method for person re-identification,jointly optimizing the retrieval and verification tasks.By substituting the nondifferentiable operations in the retrieval and verification evaluation metrics with parametrized functions,we achieve an automated design of metric learning methods based on evaluation metrics.Experimental results show that the proposed method demonstrates its superiority over other metric learning methods on public person reidentification and vehicle re-identification datasets.(3)Evaluation Metrics Combined Retrieval and Verification for Person ReidentificationTo address the problem that the evaluation aspect of person re-identification evaluation metrics is limited,this paper focuses on the absence of a specific person in the gallery set and the automatic judgment of the output results,and investigates the comprehensive evaluation metrics of multi-scenarios retrieval and verification by proposing a perfect person re-identification evaluation metrics system.Furthermore,in order to tackle the problem that it is impossible to evaluate the open-set scenario and verification mode,and the practicality of current person re-identification evaluation metrics is poor,evaluation metrics combined retrieval and verification for person reidentification is proposed to comprehensively grasp the application ranges of the algorithm.Experimental results show that the proposed evaluation metrics have better applicability for multi-scenarios re-identification research than existing evaluation metrics,whether in the retrieval or verification mode of closed-world scenario or in the verification mode of open-set scenario.In summary,this paper addresses scientific issues such as the mining of similarity relationship metric within a batch of samples,the exploration of automated metric design patterns for multi-tasks,and the discovery of comprehensive evaluation metrics for multi-scenarios.The innovative solutions are proposed for the three challenges: limited metric space of optimization strategy,limited design efficiency of metric learning method,and limited evaluation aspect of evaluation metric.The similarity relationship metric mining and automated design patterns for metric learning methods in person reidentification are effectively explored,and the diverse evaluation of person reidentification methods is investigated.The research results in both metric learning and evaluation metrics provide new insights into the application of person re-identification in practical application scenarios.