Algorithm Design And Optimization For Cross-domain Person Re-identification

With the development of society,a huge amount of cameras are deployed to monitor every corner of the cities.Intelligent data processing of these equipments has become one of the most popular research issues in academic and industrial area.Person re-identification(ReID)is getting more and more attention of many large companies and famous academic institutions because of its important application potential and significant researching meaning.The performance of ReID models has been greatly improved benefiting from the improvement of deep learning algorithms and the appearance of large-scale datasets in recent years.However,most ReID models have the same problem: the model performance will drop dramatically when the train-set and test-set come from different domains,which means the models can not get satisfactory results when they are used in another totally different scene.Focusing on the problem of weak generalization ability most ReID models face in cross-domain situation,this paper does the following research:Firstly,most ReID models mainly use the Convolutional Neural Network(such as ResNet and VGG)pretrained on image classification tasks to extract features of input images and add a fully connected layers with a softmax classifier after the output of CNN.The basic idea of these methods is to transforming the ReID task to a image classification task with the purpose of training the model by optimizing the classification loss function and every person ID is considered as a class.However,the models trained by these methods have a high demand on both the quality and quantity of training data and have weak generalizaion ability.In terms of this issue,this paper use triple loss to replace the classification loss as the optimization target in the training stage to enhance the model performance on unknown target domain.The rank1 accuracy and mAP are improved by 10% and12% respectively.Secondly,although the scale of person re-identification datasets have been increased from thousands to tens of thousands in recent years,it is still not enough for the training of deep learning model.Focusing on the problem of data deficiencies,this paper adopts two methods to extend the training dataset.The first method is using the Generative Adversarial Networks to increase the number of train-set and another method applies multi-dataset training.The experiment results show that increasing data diversity of the train-set plays a very important role in the improvement of model generalization ablity.Comparing with the most precise cross-domain person re-identification work,the method proposed by this paper are higher by 2.4% in rank1 accuracy and 14.71% in mAP.Finally,this paper also applies the model to vehicle re-identification tasks by fine tuning the model trained on person re-identification datasets with unlabeled vehicle data.The experiment results show that this method can enhance model performance on another different class re-identification problem.The fine-tuned model are higher by 14.06% in rank1 accuracy and 5.41% in mAP.