Pose Estimation-Stacked Hourglass Networks for Human Pose Estimation

0. Preface

• Relevant information:
  • arxiv
  • github(mmpose)
  • Interpretation of the paper (there are several pictures, which are clearer than the original picture of the paper)
• Basic information of the paper
  • Field: attitude estimation
  • Author unit: University of Michigan
  • Publication time: ECCV 2016
• One sentence summary: proposed an hourglass-style backbone.

1. What problem to solve

• At that time, the research of convolutional neural network in pose estimation was not enough, and various new results were still being explored to deal with pose estimation problems.
• The ultimate goal of pose estimation is to determine which pixel of the original image each key point is, so it must be a process from downsample to upsample.

2. What method was used

• The design inspiration of hourglass comes from the need to extract information of all sizes.

• The overall structure of Hourglass is shown in the figure below

  • That is, the size of the feature map keeps shrinking, increasing, shrinking, and increasing.
  • 

• For each hourglass, its structure is as follows

  • 

• There is a picture in the paper, put it down here

  • The structure of each box in each hourglass corresponds to the left picture.
  • The picture on the right introduces Intermediate Supervision.
    • How to translate, intermediate supervision? strange.
    • This means that the network is composed of multiple hourglasses, and each hourglass has to output a prediction result (that is, the key point heat map, the blue box in the figure below), and calculate the loss function.
  • 

3. How effective is it

• At that time, SOTA was achieved on both FLIC and MPII. It's not very useful to post pictures now. After all, it was a few years ago, so if you need it, you can read the paper by yourself.
• Record the training details casually
  • Single pose estimation problem
  • The image size is 256x256
  • Rotation is used for data enhancement
  • The loss function is MSE
  • GT uses a 2D Gaussian distribution

4. What are the problems & what can be learned

• It should belong to the backbone commonly used in current attitude estimation. In fact, centernet also uses hourglass as the backbone.
• However, I feel that the network is still relatively complex, it is estimated to be relatively large, and the time performance is average.