Jung Junuk ; Lee Seonhoon ; Oh Heung-Seon ; Park Yongjun ; Park Joochan ; Son Sungbin

Abstract

The goal of face recognition (FR) can be viewed as a pair similarityoptimization problem, maximizing a similarity set $\mathcal{S}^p$ over positivepairs, while minimizing similarity set $\mathcal{S}^n$ over negative pairs.Ideally, it is expected that FR models form a well-discriminative feature space(WDFS) that satisfies $\inf{\mathcal{S}^p} > \sup{\mathcal{S}^n}$. With regardto WDFS, the existing deep feature learning paradigms (i.e., metric andclassification losses) can be expressed as a unified perspective on differentpair generation (PG) strategies. Unfortunately, in the metric loss (ML), it isinfeasible to generate negative pairs taking all classes into account in eachiteration because of the limited mini-batch size. In contrast, inclassification loss (CL), it is difficult to generate extremely hard negativepairs owing to the convergence of the class weight vectors to their center.This leads to a mismatch between the two similarity distributions of thesampled pairs and all negative pairs. Thus, this paper proposes a unifiednegative pair generation (UNPG) by combining two PG strategies (i.e., MLPG andCLPG) from a unified perspective to alleviate the mismatch. UNPG introducesuseful information about negative pairs using MLPG to overcome the CLPGdeficiency. Moreover, it includes filtering the similarities of noisy negativepairs to guarantee reliable convergence and improved performance. Exhaustiveexperiments show the superiority of UNPG by achieving state-of-the-artperformance across recent loss functions on public benchmark datasets. Our codeand pretrained models are publicly available.