Data Augmentation And Domain Adaptation In Person Re-identification

With the rapid development of multimedia technology,video surveillance has been widely used in public,and person re-identification(re-ID)has become one of the popular tasks in the field of computer vision and pattern recognition.Person re-ID is an instance retrieval task,which aims to find out the same person of the query from a database collected from non-overlapped cameras.In this task,reducing the overfitting risk,learning a camera-invariant representation,learning a model that generalizes well on a new dataset and effectively leveraging unlabeled dataset are key techniques for large-scale scenarios person re-ID.This thesis presents extensive research efforts on the above key techniques,and focuses on data augmentation,image style translation,domain adaptation and re-ranking,attempting to provide theoretical guidance and practical significance for large-scale scenarios person re-ID.In the aspect of reducing the overfitting risk,we propose a simple but effective data augmentation method to augment training samples and improve the generalization ability of deep model.In the aspect of learning camera-invariant representation,we propose a camera style transfer method for improving the robustness to variations caused by cameras.In the aspect of learning generalizable model and utilizing unlabeled data,we introduce three unsupervised domain adaptation methods for improve the generalization ability of model on a new dataset.In addition,we present a re-ranking method that leverages the neighborhood relationship between unlabeled testing samples for refining the initial ranking results.The main innovations of this thesis are featured in the following five aspects.·We propose a random erasing data augmentation method.By randomly selecting a rectangle region in an image and erasing its pixels with random values,various new samples are generated for improving the generalization ability of deep model.This method is complementary to commonly used data augmentation techniques,can be integrated with various CNN models and can readily be applied to various vision tasks.·We propose a camera style transfer(CamStyle)method.We explicitly consider the camera style in person re-ID and introduce to use generative adversarial network(GAN)for camera style transferring.CamStyle can serve as a data augmentation approach that smooths the camera style disparities and thus improves the robustness to camera variations.In addition,CamStyle can be extended to unsupervised domain adaptation.·We introduce a hetero-and homogeneously learning(HHL)method for model generalization.HHL enforces two properties simultaneously,camera invariance and domain connectedness,through a triplet loss,which effectively improves the generalization ability of the model in the target testing set.·We propose an invariance learning method for unsupervised domain adaptation.We introduce an exemplar memory for effectively implementing invariance learning and present a graph based positive prediction method to facilitate the invariance learning.This approach can significantly improve the adaptation accuracy.·We propose a k-reciprocal encoding method to re-rank the re-ID results.A kreciprocal feature is calculated by encoding its reciprocal nearest neighbors into a single vector,which is used for refining similarities between images.This method could greatly improve the retrieval performance without any human interaction or any labeled data.In summary,this thesis focuses on the study of person re-ID in large-scale scenarios,and proposes various algorithms for reducing the overfitting risk,learning camera-invariant feature,unsupervised domain adaptation and effectively leveraging unlabeled samples.Extensive experimental results show the rationality and effectiveness of our proposed methods.Our methods are very practical and most of them can be easily extended to other visual tasks.