Person Re-identification Based On Hard Sample Analysis

Person re-identification(re-id)aims to match people across disjoint camera views in a multi-camera system,which can be considered as a retrieval process,and the result is presented as a ranking list.However,owing to large variations caused by pose,illumination,occlusion and viewpoint changes,the images of the same pedestrian under different cameras may be very different which cause the images of the same pedestrian are difficult to rank in the front of the ranking list.Generally speaking,the samples that are difficult to match correctly are called hard samples,and hard samples are also the biggest obstacle to the performance of the algorithm.In this thesis,from the point of view of hard samples in person re-id,we will analyze the feature distribution of hard samples and improve the performance of our algorithms.Firstly,most existing person re-id algorithms based on metric learning try to learn a global distance matrix to calculate the similarity between pedestrian images.However,due to the irregularity of feature distribution of hard samples,it is difficult for global distance matrix to match them correctly,which results in the performance degradation of the algorithm.In this thesis,we propose to combine the global metric and local information to resolve failure matching cases.The global metric is first used to calculate the similarity of the samples in test set.According to the similarity distance distribution,the hard samples suspected are screened out and the local discriminant information of these samples is mined.Then the similarity of these hard samples in test set is recalculated under the local discriminant information.Secondly,we want to directly learn a specific measurement matrix for difficult samples.We use cross-validation on the training set,obtaining the training ranking list and divide the training samples into three different hard levels.In the re-training stage,under the coarse-fine tuning mechanism,different losses are adaptively applied to samples of different hard levels with the ranking information,which can obtain a more robust metric matrix for hard samples.Applying the new metric matrix to the re-ranking task of test set can further distinguish easily confused samples and improve the accuracy of the algorithm.Finally,with the development of deep learning,the capacity of person re-id dataset becomes larger and larger.It is difficult for the model to fit a small number of hard samples in the dataset during the training process,which results in the decline of the discriminant ability of the model in the testing process.Considering the different hard level in each training batch and the different number of samples needed for training models in different training stages,we propose an adaptive hard sample mining algorithm for training a robust person re-id model.Meanwhile,an adaptive threshold of hard level can make the algorithm not only stay in step with training process harmoniously but also alleviate the under-fitting and over-fitting problem simultaneously.In addition,the designed network to implement the approach is simple and effective,and has good generalization performance which can be easily combined with existing deep learning networks and person re-id models.