行人重识别,又称行人再识别,是利用 CV 技术判断图像或视频序列中是否存在特定行人的技术。常规的行人重识别方法往往需要高昂的人工标注成本,计算复杂度也很大。在本文中,中山大学研究者提出的弱监督行人重识别方法恰恰克服了这两方面的障碍,并发布了一个大型行人重识别数据集。
论文地址:https://arxiv.org/pdf/1904.03845.pdf
代码、模型和数据集:https://github.com/wanggrun/SYSU-30k
将行人图像按拍摄时间段分组成袋并分配袋类别标签;
结合图模型和深度神经网络捕获一个袋中所有图像之间的依赖关系来为每张图像生成可靠的伪行人类别标签,作为行人重识别模型训练的监督信息;
进一步将图模型可微化,实现图模型和行人重识别模型的一体训练;
将图模型损失和重识别损失的线性组合作为总损失函数,利用反向传播算法更新网络所有层的参数。
(1)
1. 图模型行人重识别
(2)
![](https://image.jiqizhixin.com/uploads/editor/61a5de04-3d53-4503-9b52-37abfac98103/640.png)
![](https://image.jiqizhixin.com/uploads/editor/b28df9b9-fedf-4158-8155-3b1f25e8b0d8/640.png)
2. 一元项
(3)
![](https://image.jiqizhixin.com/uploads/editor/925b0ab7-ca99-47ff-9501-effef6d0de0c/640.jpeg)
![](https://image.jiqizhixin.com/uploads/editor/ed314d3c-9a31-4f1b-83ad-2c5e8256f28a/640.jpeg)
3. 成对项
(4)
![](https://image.jiqizhixin.com/uploads/editor/2e9c28b1-373d-42b9-9f2f-fd51de144ab2/640.png)
![](https://image.jiqizhixin.com/uploads/editor/ab322cfc-42c5-44b6-a62d-99acd7d827d7/640.png)
(5)
4. 袋限制
5. 伪行人类别标签的推理
![](https://image.jiqizhixin.com/uploads/editor/e44a7c6f-1e80-4ea4-949e-e4acdd7d6958/640.png)
(7)
![](https://image.jiqizhixin.com/uploads/editor/801f2bf1-84c6-4187-91b2-c832d7872690/640.jpeg)
![](https://image.jiqizhixin.com/uploads/editor/fd2f1aae-fa98-4475-8397-fe2a0ab9a6e7/640.png)
(8)
(9)
![](https://image.jiqizhixin.com/uploads/editor/2395d40c-32b5-4dd5-84e2-c14309ca5d26/640.png)
![](https://image.jiqizhixin.com/uploads/editor/c6f28510-7d72-4a93-9628-dc5b725ede82/640.png)
![](https://image.jiqizhixin.com/uploads/editor/8a490d2c-9f56-4537-ab4e-812c37fe79a9/640.png)
(10)
![](https://image.jiqizhixin.com/uploads/editor/e7bedb52-7cd7-452d-82be-1b23cc8bf406/640.png)
![](https://image.jiqizhixin.com/uploads/editor/74e1719b-a61f-474f-bda7-7378ae8e2870/640.png)
![](https://image.jiqizhixin.com/uploads/editor/9d339833-175d-44de-a70b-af69e908de4e/640.png)
![](https://image.jiqizhixin.com/uploads/editor/972cd142-ef8c-45eb-a577-d0416d56c72f/640.png)