A Hierarchical Multi-Proxy Loss with Dynamic Main-Proxy for Deep Metric Learning

Lei Zhao, Xiao-Lei Zhang

Abstract: Proxy-based approaches in deep metric learning have recieved wide interest due to their efficient training process and rapid network convergence in the past few years. However, existing single-proxy methods aim to learn a common feature representation for each class by assigning a separate proxy for each class, which contradicts the inherent intra-class variance of samples from the same class, impeding more fine-gained similarity retrieval. In this paper, we propose a hierarchical multi-proxy method named dynamic main-proxy anchor (DMA) to address this issue. The approach first assigns multiple sub-proxies to learn different intra-class features and then utilizes a dynamically constructed main-proxy to handle class-related characteristics. In addition, we propose a regularization method to ensure closeness between similar sub-proxies and distance between dissimilar ones. Experimental results on three widely-used datasets show the superiority of the proposed DMA over the state-of-the-art methods in both retrieval and clustering tasks.