1. Information
pattern recognition 2018
Lin Wu, Yang Wang, Xue Li, Junbin Gao
2. Motivation
目前很多re-id方法都是在图片上先划分出固定的区域，然后在每个区域上提取特定特征，最后根据这些特征，计算出每幅图片的差异从而得到两幅或多幅图片是否是属于同一个人。这种方法依赖于局部特征的精细程度，否则无法将正负样本予以很好地区分。
3. Method
本文提出了一种全新的深度乘积门限函数（multiplicative integration gating function）该函数可有效回答行人重识别的两个问题：what-and-where to match for effective person re-id。
具体来说：
- “what to match for effective person re-id”：两个CNN提取特征，形成特征描述子后，采用乘积门限函数的形式强化局部模式。
- “where to match for effective person re-id”：为了解决重识别中存在的行人不对齐问题，本文采用4向RNN（four-directional recurrent neural network）。
图1. 本文模型
可以看出，该模型属于孪生网络（Siamese network）。关于孪生网络的介绍，可参考如下链接：
https://blog.csdn.net/Suan2014/article/details/80599595
也是一种端到端的网络模型
4. Experiment
-
Datasets：VIPeR, CUHK03, Market-1501
-
Baselines：
(1) CNN with fully-connected layers (FC-CNN)
(2) Fisher vectors with CNN features (FV-CNN)
(3) Fisher vectors with SIFT features (FV-SIFT)
(4) Bilinear CNN with spatial pyramid pooling( BCNN+SPP)
(5) Bilinear CNN [19] with stacked four directional RNNs (B-CNN+IRNNs)
(6) Multiplication Integration networks with spatial pyramid pooling(MI+SPP)
(7)Our approach: 连体双通道CNN部分考虑三种形式：
(i) [VGG-M,VGG-M]);
(ii) [VGG-D, VGG-M]);
(iii) VGG-D, VGG-D]). -
Comparison to baselines
-
Comparison to state-of-the-art approaches
5. Conclusion
Comprehensive experiments show that our designed network achieves the superior performance on person re-identification. For the future work, we will continue to improve the models of part localization and matching with attention model.