PGDM (ICME-2018)

动机：
Existingmethods usually solve it with end-to-end multi-label deep neural networks, while the structure knowledge of pedestrian body has been little utilized. Considering that attributes have strong spatial correlations with human structures, e.g. glasses are around the head

所提方法
1.introduce pedestrian body structure into PAR
2.propose a Pose Guided DeepModel (PGDM) to improve attribute recognition
The PGDM consists of three main components:

1. coarse pose estimation which extracts the pose knowledge from a pre-trained pose estimation model
2. body parts localization which adaptively locates informative image regions with only image-level supervision
3. multiple features fusion which combines the part-based features for attribute recognition

contributions：
1.To our knowledge, this is the first attempt to explore the deformable pedestrian body structure knowledge, e.g. pose information, for pedestrian attribute recognition.
2.A pose guided deep model is proposed, which could not only transfer the pose knowledge from existing pose estimation model to surveillance scenarios, but also adaptively locate informative regions for the high-level attribute recognition task.
3.The proposed method has obtained competitive results in three large-scale pedestrian attribute datasets

主要框架：

coarse pose estimation(由Convolutional pose machines论文中得到)：
first ， adopt CPM model , which is trained on MPII and LSP with six stages, to generate 14 prior human pose key points as well as confidence scores. The human key points consist head, neck, right shoulder, right elbow, right wrist, left shoulder, left elbow, left wrist, right hip, right knee, right ankle, left hip, left knee, left ankle.
Second, taking the generated pose key points as coarse ground truth pose, we can train a regression network to regress pedestrian pose.

Adaptively Region Localization:
use Spatial Transformer Networks (STN) （论文Spatial transformer networks） for region extraction.

实验：
PETA, RAP, and PA-100K，下方三个图表按照三个数据集排序：

部分属性分别使用Main Net和PGDM的识别结果。

实验结果表明，所提出的PGDM能与全局体主网产生互补的结果，主网与PGDM的融合能产生更好的结果