期刊
KNOWLEDGE-BASED SYSTEMS
卷 203, 期 -, 页码 -出版社
ELSEVIER
DOI: 10.1016/j.knosys.2020.106154
关键词
Person re-identification; Attention module; Score learning head
资金
- National Natural Science Foundation of China [61802111]
- Foundation of Henan Education Department, China [19A520002]
- Postdoctoral Science Fund of China [2015M582182]
Person re-identification (re-ID) refers to matching people across multiple camera views at different times and locations. The challenge is mainly about the huge variance of visual appearance of a specific pedestrian owing to pose variations, illumination changes and various camera-styles. In this paper, an Attention-Aware Scoring Learning (AASL) framework is proposed to address these issues. The proposed AASL framework consists of two attention modules and a score learning head. Specifically, the two modules, Spatial Attention Grid and Channel Attention Grid, embedded respectively in the shallow and deep layer in the convolutional neural network, are put forward to help the network learn the most discriminative visual features. Furthermore, an adaptive module termed score learning head is proposed to optimize the parameters of the attention modules. The present paper carries out extensive experiments on three large-scale datasets, including Market-1501, DukeMTMC-reID and CUHK03, after which it is found that our Attention-Aware Scoring Learning framework significantly outperforms the baseline model and achieves a competitive performance compared with the state-of-the-art person re-ID methods. (C) 2020 Elsevier B.V. All rights reserved.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据