Learning Robust Embedding Representation With Hybrid Loss for Classification and Verification

Abstract

This paper presents a method of building an embedding representation via deep metric learning, which works well in both classification and verification problems. The embedding is built via a proposed hybrid loss, which consists of a softmax loss and a Euclidean-metric loss. The hybrid loss explores the tradeoff, balancing the discriminativeness and invariance. In deep metric learning, a softmax loss is proposed for classification and a Euclidean-metric loss is responsible for verification. We apply the proposed loss and the softmax loss to the well-known deep models, such as VGG-16 and ResNet-50, respectively. In addition, the performances are evaluated by two datasets (CIFAR and Market1501), in which the CIFAR is for classification and the Market1501 is for verification. The results indicate that the embedding learned from the proposed loss does improve the performance on both the classification and verification tasks.

Keywords

• Embedding
• metric learning
• classification
• verification
• person re-identification