Learning comprehensive global features in person re-identification: Ensuring discriminativeness of more local regions

• Published in: Pattern recognition Vol. 134; p. 109068
• Main Authors: Xi, Jiali, Huang, Jianqiang, Zheng, Shibao, Zhou, Qin, Schiele, Bernt, Hua, Xian-Sheng, Sun, Qianru
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.02.2023
• Subjects: Baseline; Comprehensive; Person re-identification; Baseline; Comprehensive; Person re-identification
• ISSN: 0031-3203; 1873-5142
• DOI: 10.1016/j.patcog.2022.109068

•A novel baseline for person re-identification is proposed to learn comprehensive global embedding, ensuring that more local regions (the number of local regions is manually defined) of global feature maps are discriminative.•A Non-parameterized Local Classifier (NLC) module is designed to obtain a score vector of each local region on feature maps in a non-parametric manner.•A Comprehensive Global Embedding (CGE) module is designed to revise the global logits such that the subsequent cross entropy loss up-weights the loss assigned to samples with hard-to-learn local regions.•The network achieves 65.9% mAP, 85.1% rank1 on MSMT17, 86.4% mAP, 87.4% rank1 on CUHK03 labeled, 84.2% mAP, 85.9% rank1 on CUHK03 detected, and 92.2% mAP, 96.3% rank1 on Market-1501. Person re-identification (Re-ID) aims to retrieve person images from a large gallery given a query image of a person of interest. Global information and fine-grained local features are both essential for the representation. However, global embedding learned by naive classification model tends to be trapped in the most discriminative local region, leading to poor evaluation performance. To address the issue, we propose a novel baseline network that learns strong global feature termed as Comprehensive Global Embedding (CGE), ensuring more local regions of global feature maps to be discriminative. In this work, two key modules are proposed including Non-parameterized Local Classifier (NLC) and Global Logits Revise (GLR). The NLC is designed to obtain a score vector of each local region on feature maps in a non-parametric manner. The GLR module directly revises the logits such that the subsequent cross entropy loss up-weights the loss assigned to samples with hard-to-learn local regions. The convergence of the deep model indicates more local regions (the number of local regions is manually defined) on the feature maps of each sample are discriminative. We implement these two modules on two strong baseline methods including the BagTricks (BOT) [1] and AGW [2]. The network achieves 65.9% mAP, 85.1% rank1 on MSMT17, 86.4% mAP, 87.4% rank1 on CUHK03 labeled, 84.2% mAP, 85.9% rank1 on CUHK03 detected, and 92.2% mAP, 96.3% rank1 on Market-1501. The results show that the proposed baseline achieves a new state-of-the-art when using only global embedding during inference without any re-ranking technique.