Page 361 - 《软件学报》2021年第12期

软件学报 ISSN 1000-9825, CODEN RUXUEW E-mail: jos@iscas.ac.cn
Journal of Software,2021,32(12):4025−4035 [doi: 10.13328/j.cnki.jos.006108] http://www.jos.org.cn
©中国科学院软件研究所版权所有. Tel: +86-10-62562563

∗
近邻中心迭代策略的单标注视频行人重识别

1
1
1
1
1
1
张云鹏 , 王洪元 , 张 继 , 陈 莉 , 吴琳钰 , 顾嘉晖 , 陈 强 2
1
(常州大学 信息科学与工程学院,江苏 常州 213614)
2 (社会安全信息感知与系统工业和信息化部重点实验室(南京理工大学),江苏 南京 210094)
通讯作者: 王洪元, E-mail: hywang@cczu.edu.cn

摘 要: 为解决视频行人重识别数据集标注困难的问题,提出了基于单标注样本视频行人重识别的近邻中心迭代
策略.该策略逐步利用伪标签视频片段迭代更新网络结构,以获得最佳的模型.针对预测无标签视频片段的伪标签准
确率低的问题,提出了一种标签评估方法:每次训练后,将所选取的伪标签视频片段和有标签视频片段特征中每个类
的中心点作为下一次训练中预测伪标签的度量中心点;同时提出基于交叉熵损失和在线实例匹配损失的损失控制
策略,使得训练过程更加稳定,无标签数据的伪标签预测准确率更高.在 MARS,DukeMTMC-VideoReID 这两个大型
数据集上的实验验证了该方法相比于最新的先进方法,在性能上得到非常好的提升.
关键词: 视频行人重识别;近邻中心迭代策略;标签评估方法;单标注;损失控制策略
中图法分类号: TP391

中文引用格式: 张云鹏,王洪元,张继,陈莉,吴琳钰,顾嘉晖,陈强.近邻中心迭代策略的单标注视频行人重识别.软件学报,2021,
32(12):4025−4035. http://www.jos.org.cn/1000-9825/6108.htm
英文引用格式: Zhang YP, Wang HY, Zhang J, Chen L, Wu LY, Gu JH, Chen Q. One-shot video-based person re-identification
based on neighborhood center iteration strategy. Ruan Jian Xue Bao/Journal of Software, 2021,32(12):4025−4035 (in Chinese).
http://www.jos.org.cn/1000-9825/6108.htm
One-shot Video-based Person Re-identification Based on Neighborhood Center Iteration Strategy

1
1
1
1
1
1
ZHANG Yun-Peng , WANG Hong-Yuan , ZHANG Ji , CHEN Li , WU Lin-Yu , GU Jia-Hui , CHEN Qiang 2
1
(School of Information Science and Engineering, Changzhou University, Changzhou 213614, China)
2
(Key Laboratory of Information Perception and Systems for Public Security of MIIT (Nanjing University of Science and Technology),
Nanjing 210094, China)
Abstract: In order to solve the problem of labeling difficulty in video-based person re-identification dataset, a neighborhood center
iteration strategy based on one-shot video-based person re-identification is proposed in this study, which gradually optimizes the network
by using pseudo-labeled tracklets to obtain the best model. Aiming at the problem that the accuracy of predicting pseudo labels of
unlabeled tracklets is low, a novel label evaluation method is proposed. After each training, the center points of each class in the features
of the selected pseudo-labeled tracklets and labeled tracklets are used as the measurement center points for predicting the pseudo labels in
the next training. At the same time, a loss control strategy based on cross entropy loss and online instance matching loss is proposed in
this study, which makes the training process more stable and the accuracy of the pseudo labels higher. Experiments are implemented on
two large datasets: MARS and DukeMTMC-VideoReID, and the result demonstrates that the proposed method outperforms the current
state-of-the-art methods.

∗ 基金项目: 国家自然科学基金(61976028, 61572085, 61806026, 61502058); 江苏省自然科学基金(BK20180956); 社会安全信息
感知与系统工业和信息化部重点实验室(南京理工大学)创新基金(202004)
Foundation item: National Natural Science Foundation of China (61976028, 61572085, 61806026, 61502058); Natural Science
Foundation of Jiangsu Province (BK20180956); Key Laboratory Foundation of Information Perception and Systems for Public Security of
MIIT (Nanjing University of Science and Technology) (202004)
收稿时间: 2020-01-15; 修改时间: 2020-04-19; 采用时间: 2020-05-21; jos 在线出版时间: 2020-10-12