Deep Fusion Feature Presentations for Nonaligned Person Re-Identification 论文简要总结

Digital Object Identifier 10.1 109/ACCESS.2019.2920426

论文题目:用于非对齐行人重新识别的深度融合特征表示

论文链接: https://www.researchgate.net/publication/333590933_Deep_Fusion_Feature_Presentations_for_Nonaligned_Person_Re-identification.
代码链接: https://github.com/Henuzhaoyli/Twostream_reID.

1 摘要

目前行人冲识别的研究面临着各种挑战。相似的外表,相机角度,场景光照度,行人姿态等。
作者提出了一个新的双流网络,称为空间分割网络,它在一个统一的框架内学习全局和局部特征,用于非对齐行人的重识别。一个流分支集中在深度卷积神经网络中使用全局自适应平均汇集的空间特征学习。另一种是利用水平条纹平均池化来学习细微的局部特征。

2 理论研究

Deep Fusion Feature Presentations for Nonaligned Person Re-Identification 论文简要总结
backbone 是基于resnet 50,第二个分支(和PCB一样)

3 损失函数

基于resnet50的主干网络,利用的损失函数为交叉熵损失函数,最后的距离度量用的是余弦距离

Deep Fusion Feature Presentations for Nonaligned Person Re-Identification 论文简要总结
Deep Fusion Feature Presentations for Nonaligned Person Re-Identification 论文简要总结

4 实验结果

Deep Fusion Feature Presentations for Nonaligned Person Re-Identification 论文简要总结
Deep Fusion Feature Presentations for Nonaligned Person Re-Identification 论文简要总结
Deep Fusion Feature Presentations for Nonaligned Person Re-Identification 论文简要总结

re-ranking方法Z. Zhong, L. Zheng, D. Cao, and S. Li, ‘‘Re-ranking personre-identification with κ-reciprocal encoding,’’ in Proc. IEEE Conf.Comput. Vis. Pattern Recognit., Jul. 2017, pp. 1318–1327.

re-ranking链接: https://arxiv.org/abs/1701.08398.