Multi-Camera Transfer GAN For Person Re-Identification

Person re-identification is a new technology emerging in the field of computer vision in recent years.It is an important task in many security and monitoring applications and has attracted more and more attention in computer vision.Person re-identification refers to the retrieval of target pedestrians in non-overlapping lens.Scenes and angles of multiple cameras are completely different.Suppose we want to identify a pedestrian in one camera,that is,to retrieve the pedestrian in other cameras.In addition to the difference in appearance of the target pedestrian itself under different cameras,the influence of other pedestrians should also be considered.For example,the target pedestrian needs to be compared with multiple pedestrians in other cameras.With the development of generative adversarial networks and related algorithms,person re-identification based on generative adversarial networks has made good progress.The model training using the image generated by the generative adversarial networks can improve the performance.At present,person re-identification performance in a single dataset has been significantly improved,but person re-identification model trained in one dataset often can't work well in another dataset.The main work of this paper is as follows:(1)Person re-identification is also a sub problem of image retrieval,so the idea of fine-grained image retrieval is introduced in this paper.In this paper,the MSCDA(Mixed Selective Convolutional Descriptor Aggregation)method is proposed,which combines the SCDA(Selective Convolutional Descriptor Aggregation)features with the global average pooling features,then obtain the final pedestrian features.The method can select useful deep descriptors,at the same time,remove background by localizing the main object.(2)We propose CTGAN(Multi-Camera Transfer GAN),an image-to-image translation method.In this method,image to image translation can be performed on multiple lens fields of pedestrian dataset using only a single model.The marked training image is transferred to each camera of the target dataset.At the same time,for the feature learning model,we adopt the selective convolution descriptor aggregation method,which can locate the main pedestrian object in the image,filter out the background noise,and keep the useful depth descriptor,to improve the retrieval accuracy.