0. Summary
In actual solution, the estimation of human body posture is often transformed into the prediction problem of key points of the human body, that is, first predicting the position coordinates of each key point of the human body , and then determining the spatial position relationship between the key points based on prior knowledge , thus obtaining Predicted human skeleton.
Pose estimation problems can be divided into two major categories: 2D pose estimation and 3D pose estimation. As the name suggests, the former predicts a two-dimensional coordinate (x, y) for each key point; the latter predicts a three-dimensional coordinate (x, y, z) for each key point, adding one-dimensional depth information.
1. 2D pose estimation
For 2D pose estimation, most of the current research is multi-person pose estimation, that is, each picture may contain multiple people. There are usually two ways to solve this type of problem: top-down and bottom-up:
The idea of top-down is to first perform target detection on the image and find all the people; then crop the people out of the original image, resize them and input them into the network for pose estimation. In other words, top-down transforms the problem of multi-person pose estimation into the problem of multiple single-person pose estimation.
The idea of bottom-up is to first find all the key points in the picture, and then group the key points to get an individual.
1.1Top-down
The input of the network is the bounding box containing a person, and
the output of the network is the coordinates of k key points of the person. Because there are horizontal and vertical axes, a total of 2k numbers need to be returned. There are two prediction ideas:
1. Directly regress the coordinates: the output of the network is 2k numbers output through the fc layer 2.
heatmap: predict the heat value of each point of the feature map to obtain the heat map, and the maximum value on the heat map is Coordinates of key points
Reference:
https://zhuanlan.zhihu.com/p/104917833