Research On Person Re-identification Based On Neural Architecture Search

Person re-identification is one of the core technologies in the field of intelligent video surveillance.In recent years,person re-identification methods based on convolutional neural network have made breakthrough progress.However,most of the existing methods are built on classic classification backbone networks pre-trained with external data.These network architectures rely on manual design,the architecture is relatively fixed,and rely heavily on the pre-training of external proxy data.Although there are some methods to design networks specifically for person re-identification task,manually designing the network strongly depends on expert experience,which is inefficient and difficult to achieve optimal settings.In response to the above problems,this paper has conducted an in-depth research on the person re-identification based on Neural Architecture Search(NAS),the main contributions are as follows:(1)Person re-identification architecture search method based on retrieval objective is proposed.NAS methods of automatic design of optimal network architecture are basically proposed to solve classification tasks,but using the classification accuracy to guide the direction of the architecture search is not suitable for person re-identification.In order to coordinate the inconsistency between the two optimization goals,this paper proposes person re-identification architecture search method based on retrieval objective,which establishes a new retrieval objective with the assistance of triplet loss to correctly guide the direction of architecture search,not only the classification accuracy is considered,but also the distance between different samples is used.The ranking performance of the algorithm is improved by maximizing the feature discrimination ability between different pedestrians.Experiments show that without relying on external data,the architecture search method is used to automatically find a high-performance network architecture that is particularly suitable for pedestrian re-identification tasks on the target data set,which can extract discriminative pedestrian features for similarity matching.(2)Person re-identification architecture search method based on multi-granularity features is proposed.Considering that the current NAS algorithm rarely uses pedestrian-specific body structure information,in order to further improve the performance of the search algorithm,based on the retrieval objective,this paper proposes a person re-identification architecture search method based on multi-granularity features.The granularity features of multiple pedestrian body-parts are introduced into the search algorithm,and cooperate with the twolevel pedestrian overall features.The multi-objective loss with global coarse-grained features combined with local fine-grained features optimize the network architecture space and encourages search algorithms to learn more detailed information to guide the direction of architecture search.Experimental results show that the architecture search method automatically generates a better person re-identification network and achieves better performance than most manually designed network architectures.