摘要
在行人重识别模型中引入邻域数据关系,提出了一种基于图像邻域相似度的重排序方法。首先扩充图像的邻域数据,然后计算图像对不同邻域数据的相似度权重,利用该权重得到代表邻域相似度的分布距离,再用分布距离与原始距离计算得出最终距离作为重排序评判标准。使用CCL,Transreid, Torchreid等行人重识别模型在Market-1501,DukeMTMC-reID数据集上进行实验,结果表明本文方法对基准模型的精度提升均超过该领域的主流算法,证实了本文方法的有效性和泛化性。该重排序方法不需要任何人工交互和额外数据,适用于大规模数据集,可以有效应用于图像检索、目标跟踪等需要考虑相似度关系的任务中。
The neighborhood data relationship was introduced into person re-identification model,and a re-ranking method based on image neighborhood similarity was proposed.Firstly,the neighborhood data of the image was expanded,and then the similarity weights of the image to different neighborhood data were calculated,and the distribution distance representing the neighborhood similarity was obtained by using the weight.Finally,the final distance which was calculated by using the distribution distance and the original distance was taken as the re-ranking criterion.CCL,Transreid,Torchreid and other person re-identificationmodels were used to conduct experiments on Market-1501 and DukeMTMC-reID datasets.The results show that the accuracy improvement of the baseline model by the proposed method exceeds that of the mainstream algorithms in the field,which proves the effectiveness and generalization of the proposed method.The re-ranking method does not require any human interaction and additional data.It is suitable for largescale datasets and can be effectively applied to tasks that need to consider the similarity relationship such as image retrieval and object tracking.
作者
吕翔
陈念年
蒋勇
LV Xiang;CHEN Niannian;JIANG Yong(School of Computer Science and Technology,Southwest University of Science and Technology,Mianyang 621010,Sichuan,China)
出处
《西南科技大学学报》
CAS
2023年第4期96-103,共8页
Journal of Southwest University of Science and Technology
基金
四川省科技厅重点研发项目(2021YFG0031)
四川省省级科研院所科技成果转化项目(22YSZH0021)。
关键词
行人重识别
重排序
邻域相似度
Person re-identification
Re-ranking
Neighborhood similarity
