Title: Person Re-Identification Using CNN Features Learned from Combination of Attributes
Authors: Tetsu Matsukawa, Einoshin Suzuki
Affliation: Faculty of Information Science and Electrical Engineering (ISEE), Kyushu University, Japan(日本九州大学)
* Papers about "Person Re-Identification" via deep learning methods 行人重识别深度学习方法相关论文整理
Contribution
- We show that the fine-tuning on the pedestrian attribute dataset largely improves the performance of CNN features for person re-identification.
- We propose a loss function for classifying combination attributes to increase discriminative power of CNN features.
论文概括
基本思想:利用行人的attributes增强对行人识别的准确性
主要方法:
- 基本网络结构是AlexNet
- Phase 1:fine-tune,利用Pedestrian Attribute (PETA) dataset
- task 1: 利用Pedestrian Attribute (PETA) dataset训练行人的属性,如年龄、性别 、Luggage(是否带行李)、上体衣服颜色等等,每个属性分成一个group,共G个。该部分构建Multi-Attribute Loss,该部分网络的FC8层有G个,每个group对应一个softmax分类。
- task 2: 构建Combination-Attribute Loss,将所有group的所有类别放在一起训练,更注重全局性。
- Phase 2:利用学习到的feature应用在行人重识别数据集上。主要是对上述训练网络的FC6特征进行输出,再进行metric learning(此处利用CVPR15年已有的方法,待阅读笔记)