摘要
针对行人重识别特征提取过程中特征图分辨率不断下降,丢失大量空间信息和细节信息,导致特征鲁棒性较低的问题,提出一种基于高分辨率特征提取网络的行人重识别方法。采取变换背景的方法对训练数据集进行数据扩充,提高数据样本的多样性;通过构建高分辨率特征提取网络,使得在整个特征提取过程中网络里始终拥有高分辨特征;结合三元损失函数和改进的交叉熵损失函数进行网络的训练。该高分辨率特征提取网络行人重识别方法在Market1501数据集上Rank-1达到了94.6%,mAP为86.0%。在DukeMTMC-reID数据集上Rank-1达到了90.3%,mAP为78.2%。该方法在两大数据集上均有良好表现,具有一定的应用价值。
The resolution of feature maps constantly decreases and a large amount of spatial and detail information is lost during the feature extraction process of pedestrian re-identification,which leads to low robustness.To solve this problem,we proposed a pedestrian re-identification method based on high-resolution feature extraction network.The training data set was expanded by changing the background of images to increase the diversity of data samples.We constructed a high-resolution feature extraction network,which had high-resolution features in the entire feature extraction process.The triplet loss and the improved cross-entropy loss function was adopted to train the network.The proposed method has Rank-1 of 94.6%and mAP of 86.0%on the Market1501 data set.On the DukeMTMC-reID data set,Rank-1 reaches 90.3%and mAP is 78.2%.The method performs well on two large data sets and has certain application value.
作者
董明超
黄伟
吴金明
徐怀宇
Dong Mingchao;Huang Wei;Wu Jinming;Xu Huaiyu(School of Information Science and Technology,ShanghaiTech University,Shanghai 200120,China;Shanghai Advanced Research Institute,Chinese Academy of Sciences,Shanghai 200120,China;University of Chinese Academy of Sciences,Beijing 100049,China)
出处
《计算机应用与软件》
北大核心
2022年第3期180-186,共7页
Computer Applications and Software
基金
中国科学院战略性先导科技专项(XDC02000000,XDC02070700)。
关键词
行人重识别
高分辨率
特征提取
数据增强
Pedestrian re-identification
High-resolution
Feature extraction
Data augmentation
