快速人体姿态估计--Pose Proposal Networks

Pose Proposal Networks
ECCV2018

本文使用 YOLO + bottom-up greedy parsing 进行人体姿态估计

its total runtime using a GeForce GTX1080Ti card reaches up to 5.6 ms (180 FPS)

人体姿态估计总的来说有两大类方法： top-down and bottom-up
top-down： 就是首先检测图像中的所有人，然后分别对每个人进行人体姿态估计
one detects person instances first and then applies single-person pose estimators to each detection

bottom-up：首先提取出图像中所有的人体部件 person parts,然后 对部件进行聚类，讲属于同一个人的部件连接起来。
detects parts first and then parses them into each person instance

从时间效率的角度来说， bottom-up 更具优势，它的时间不会随着图像人数的增加而线性增加

Human pose detection is achieved via the following steps.
1. Resize an input image to the input size of the CNN.
2. Run forward propagation of the CNN and obtain RPs of person instances and parts
and limb detections.
3. Perform non-maximum suppression (NMS) for these RPs.
4. Parse the merged RPs into individual people and generate pose proposals.

面对拥挤人群的姿态就力不从心了，网格检测的弊端啊

11