| With the development of society and the improvement of public security awareness,monitoring probes appear in every corner of life,which effectively prevent a series of violent and illegal events,but even so,there are still many problems in the cross camera pursuit of suspects by public security investigators.In order to solve the problem of person re-identification(re-ID)and retrieval in cross camera scenes,person re-ID emerges as the times require and becomes a research hotspot in the field of computer vision.The advantage of person re-ID is that it can lock a specific pedestrian without obtaining a clear face image and realize continuous tracking across cameras.In the research of re-ID,the most common method is supervised learning.Supervised learning needs to rely on complete data markers,and the cost of labeling the data of people in each training set under different cameras is very high.Unsupervised learning does not need data annotation,but the experimental results are unsatisfactory due to the lack of guidance of labeled data.Between supervised learning and unsupervised learning is weakly supervised learning.Therefore,weakly supervised learning has become a research hotspot,which has the advantages of both supervised learning and unsupervised learning.It can make full use of both effective supervision information and unlabeled data.In the specific CNN model training method with weakly supervision,the most commonly used are style-transferred images and one-example.style-transferred images refers to the style exchange between two datasets,which makes the original dataset A have the style of dataset B,The one-example labeling means that each person in the dataset has only one labeled sample and multiple unlabeled samples.The specific method is to use a small amount of labeled data to train a CNN model,then use the trained model to predict the label(pseudo-label)for unlabeled data,and finally use the predicted pseudo label data and the original small amount of label data to retrain the model.But because the angle,color difference and background of the pictures under different cameras are different,the pictures taken by the same pedestrian are very different.When only one sample is labeled,there is no image for cross camera learning,whichleads to low recognition efficiency.In order to solve the above problems,this paper proposes the following methods:(1)In this paper,a new One-Example annotation scheme is proposed,that is,through cycleGAN,the image styles of different cameras are transformed to each other,so as to achieve the goal that each person across the camera has at least one labeled image under each camera style.Each person has a labeled image in all different camera styles.The purpose of this is to effectively solve the problem that the One-Example labeled image is not representative.(2)In this paper,the CNN model is initialized by the single annotation sample of the original dataset and the single annotation sample of the camera-style-transferred dataset.The CNN model which only uses the data set of single style camera style for initialization has low recognition performance.The robustness of CNN model will be enhanced by using the real image of each person and the cross-camera-style-transferred image to initialize together.In the training process,using transferred image data and original data to train together greatly improves the re ID performance.In the aspect of finding trusted images to be labeled,this paper proposes to calculate the image characteristics of the original dataset,randomly select a part of the images generated from the original image and camera-style-transferred to be labeled and put into the iterative training model.A large number of experiments show the superiority of the proposed method in weakly supervised for person re-ID. |
