Unsupervised And One-shot Person Re-identification Algorithms

Person re-identification refers to the process of judging whether a specific pedestrian exists in an image or video through the technology of computer vision.However,since traditional person re-identification requires training samples to be acquired across multiple cameras,the inherent differences between the cameras cause the manual labeling method to be inefficient and cannot be applied to large-scale datasets,which limits the further development of person re-identification research.As a result,more and more researchers are beginning to focus on person re-identification studies that do not require manual labeling(i.e.,unsupervised)or require only one labeled image(single sample)per pedestrian.This paper will also propose two different algorithms for these two tasks to solve the problem of training the person re-identification model under insufficient pedestrian image label information(unsupervised and single sample).For these two types of tasks,due to the lack of category label information and pairwise label information in pedestrian images,we cannot directly perform training based on classification loss and triplet loss.Therefore,both need to estimate the label information of unlabeled pedestrian images for model training,and are usually estimated by the feature similarity between pedestrian images.However,factors such as shooting angles of different cameras,different poses of pedestrians,and changes in light cause large differences in the images of different cameras of the same pedestrian(intra-class differences),which are even greater than image differences between different pedestrians(inter-class differences).This is also one of the biggest challenges for person re-identification.In the absence of pedestrian image label information,the inherent problems of large intra-class differences and small inter-class differences will greatly affect the method of estimating labels based on the feature similarity of pedestrian images.Therefore,we propose two training frameworks under unsupervised and one-shot task settings to alleviate this problem.In the former,clustering is used to simultaneously estimate category labels and pairwise labels for multi-loss joint training,where based on pairwise labels,it is proposed to simultaneously mine trusted and difficult samples to mitigate the effects of intra-class and inter-class differences.In the latter,we fuse feature matching labels and classification prediction labels to estimate the labels of pedestrian images,and use cross-camera image hybrid generation to mitigate the effects of intra-class differences.Next,we will introduce the specific methods under these two settings.Aiming at the problem of unsupervised person re-identification,we propose an unsupervised learning algorithm based on clustering guidance,which aims to fully mine and use category label information and paired label information in unlabeled images.Specifically,we cluster the unlabeled target dataset and the labeled auxiliary dataset together.On the one hand,the information generated after the clustering of samples belonging to a certain cluster is used as cluster-level classification label information,and a non-parameterized softmax classification loss is proposed;On the other hand,based on labeled auxiliary images in the cluster as a reference point to mine the pairwise label information of the unlabeled target image,we propose the reliable sample and hard sample mining and the corresponding weighted P2 S triple loss.Therefore,we obtain two kinds of label information of unlabeled images through clustering at the same time,and jointly train the models corresponding to the two proposed loss functions.Aiming at the problem of one-shot person re-identification,we propose a person re-identification method based on double-layer pseudo-label fusion and cross-camera image mixing.Double-layer pseudolabel fusion refers to combining feature matching labels and classification prediction labels at the same time.The former refers to pseudo-labels obtained by comparing the feature similarity of the unlabeled image with the prototype of each pedestrian category.The latter refers to the prediction class probability of the model for unlabeled images.We combine the two at the same time to estimate the pseudo labels of unlabeled images,and use the information entropy criterion to progressively select the unlabeled images with the lowest information entropy for labeling.Cross-camera image hybrid generation refers to selecting the real label image and the false label image of the same pedestrian under different cameras to perform pixel-level linear interpolation to generate a pedestrian fusion image.Such a fused image can alleviate the problem of large differences in the images of the same pedestrian,and is conducive to the smooth training and learning of the network.