Proxy-AN loss for deep metric learning.

Abstract

In deep metric learning, proxy-based losses aim to introduce proxy representations to approximate class distributions, reducing training complexity and accelerating model convergence. Existing proxy-based losses can be broadly categorized into two types: sample-centric losses, which center on samples to pull positive proxies closer and push negative proxies away; and proxy-centric losses, which reverse the roles of samples and proxies. Despite their respective contributions, these methods often focus excessively on either samples or proxies, resulting in a neglect of proxy fidelity in sample-centric losses and a lack of sample discriminability in proxy-centric losses. This imbalance may lead to a sub-optimal embedding space. In this work, we introduce the Proxy-Anchor-Negative (Proxy-AN) loss, which reconciles these divergent focuses and combines the advantages of both sample-centric and proxy-centric losses. For positive pairs, we employ a proxy-centric approach to improve intra-class compactness by closely aligning anchor proxies with positive samples. For negative pairs, we utilize a sample-centric approach to enhance sample discriminability by distancing samples from all negative proxies. This strategy ensures a holistic and synergistic enhancement in the representation quality of both proxies and samples, facilitating the learning of discriminative metrics. Our extensive experiments on mainstream image retrieval benchmark datasets demonstrate substantial improvements over current leading metric learning algorithms. We also evaluate our method under partial training data and class imbalance data settings, further highlighting its superior performance across diverse scenarios.