ABSTRACT
As deep learning methods achieve a great success in computer vision, it is also expected that learning deep architectures is potentially useful for person reidentification. However, because of large variations of pedestrian data (e.g., in poses, illumination, occlusions, camera views, etc.), it is difficult to train a practical deep neural network. To deal with this issue, it is critical to find an effective way to embed the pedestrian data from irregular distributions to an appropriate feature space and learn discriminative metrics via deep convolutional networks. In this chapter, we introduce two representative methods of learning deep embeddings and deep metrics for person reidentification. The first method employs the Siamese network architecture for the first time in person reidentification, and accomplishes the deep metric learning through the cosine similarity and the loss of binomial deviance. The second method achieves robust and stable training with a particular training sample mining strategy, and implements the Mahalanobis distance in the network with a weight constraint. Because of these improvements of network structures and training methods, the deep neural network demonstrates effective feature embedding and robust metric for person reidentification.
