Proxy-based losses and pair-based losses for face image retrieval

Distance metric learning has been considered an effective strategy to represent data in computer vision problems such as image retrieval and face verification. Metric learning attempts to minimize a loss function in order to transform data into a more optimal representation for further applications. In this paper, we compare 4 different types of loss functions, e.g. 2 pair-based losses (Contrastive loss and Triplet Margin Ranking loss), and 2 proxy-based losses (Proxy-NCA loss and Proxy-Anchor loss) in a multi-class classification task. Our experiments show that the Proxy-Anchor loss could achieve 70.8% accuracy on average compared to the Proxy-NCA loss, Triplet Margin Ranking loss and Contrastive loss which could only achieve 65.5%, 62.2%, and 36.6% respectively. Furthermore, we also present the qualitative results using high dimensional plot visualization in order to evaluate data distribution and sample image retrieval results. Overall, the Proxy-Anchor loss performs better than the other losses in terms of accuracy, recall, data separation, and image retrieval.