Person Re-identification Based On Still Image

Person Re-identification is an important vision task,aiming to recognize the same person across different camera views.It has played an increasingly important role in the modern visual surveillance system,which underpins many crucial applications such as searching for criminal and tracking individuals in crowds.Currently,in computer vision,Person Re-identification is a very challenging task due to the uncertainties of illumination condition,pose of person,camera view and image quality.To improve the performance of Person Re-identification,this paper designs four algorithms from four different aspects:1)To consider the importance of different pair,this paper designs the adaptive verification loss.It merge the Cross-Entropy and Triplet Loss,forcing the network to learn more discriminate information.Particularly,the innovation is divided into two aspects: on the one hand,we replace the Euclidian Distance with the probability after Soft Max.On the other hand,we design a method to measure the importance of the triplet,furthremore giving the weight to the triplet according to its importance to make the triplet treated differently.2)To address the mismatch problem,this paper designs the locality matching loss.it is based on the DTW(Dynamic Time Warping)and is merged into triplet loss.Using this loss,the similarity of two images is not only based on the global feature matching,instead,it matches the most similar patch of the images.Particularly,the image is split into several stripes along the height direction.Then the stripes match other stripes making using of DTW.3)To make the feature containing more information,this paper designs a network which can merge the low-level inputs and high-level inputs.To merge the low-level detail information and high-level discriminate information and make the final feature is integrated with detail and discriminant,we put the low-level features and high-level features into a ”time sequence”,as the input of GRU,the final output of GRU is the final feature.In this way,the low-level features and high-level features can be integrated directly without long convolutional layer.4)To overcome the overfitting problem,this paper designs a network to make the feature keep the image original information and final discriminate information.To make the final feature is integrated with original image information and strong discriminate ability,avoiding the overfitting,we design a new network constraint.Particularly,based on the original network,we add a sub-network,which get the final features as input to reconstruct the image and compare with the original image,the similarity measure method is structural similarity(SIMM).In summary,To improve the performance of Person Re-identification deep net,we propose four methods from sample mining,network architecture,feature fusion and training procedure.