AEA-Net:Affinity-supervised entanglement attentive network for person re-identification

Abstract

Most existing person re-identification algorithms prioritize the extraction of effective and salient local features while neglecting the affinity between local and adjacent features. This phenomenon will cause misidentification when different people have similar local features. To address this issue, we introduce the AEA-Net, which emphasizes the affinity between the local features of a single image. Specifically, three important components are proposed. The affinity-supervised attention module (ASA) centers on adjacent features and utilizes the affinity between global and adjacent features to supervise the learning of attention. The affinity relationship module (AR) focuses on constructing relationship features between local and adjacent features to enhance the closeness between local features. The tangle hybrid loss (THL) makes the final predictions have a distinct weight profile. Extensive experiments quantitatively and qualitatively demonstrate that our method outperforms the state-of-the-art approaches.