Siamese cosine network embedding for person re-identification

Abstract

In person re-identification, feature embedding is the key point for new coming identities. Most state-of-the-art models adopt the features learned by convolutional neural networks (CNNs) to do similarity comparison. However, the learned features are not good enough for new identities because CNNs are designed for classification of class-known objects, not for similarity comparison of any two identities. To improve feature embedding, we propose a pairwise cosine loss based on cosine similarity measurement. Subsequently, we design a Siamese cosine network embedding (SCNE) to learn deep features for person re-identification. It is based on the Siamese architecture, with intra-class input pairs and joint supervision by the softmax loss and the pairwise cosine loss. Experimental results show that our SCNE achieves the state-of-the-art performance on the public Market1501 and CUHK03 person re-ID benchmarks.