2D human posture estimation deeppose

【https://github.com/Naman-ntc/Pytorch-Human-Pose-Estimation】 

【CVPR】【HTML】

Some early top-down deep learning methods used neural networks to directly predict the location of key points of human body 2D coordinate .

DeepPose [1] It is a classic representative of this kind of method .DeepPose Cascade neural network is used to predict the relative coordinates of each key point of human body . Each stage takes the output coordinates of the previous stage as the input , And further predict more accurate coordinate positions . Final , Convert the predicted normalized relative coordinates into absolute coordinates .

Alexander Toshev and Christian Szegedy Proposed DeepPose The earliest will be CNN（ Convolutional neural networks ） Applied to human joint point detection .DeepPose Human posture estimation is transformed into joint point regression problem , And put forward that CNN Method applied to human joint point regression ： Use the whole image input to 7 layer CNN To do joint point regression , Further more , Use cascaded CNN Detector to increase the accuracy of joint point positioning .

 DeepPose stay LSP On dataset [email protected] Average accuracy reaches 61%, It was at the time state-of-art Method

The method of direct regression of coordinates , You can get the key position directly , There is often a faster prediction speed . However , Due to the great degree of freedom of human posture , The modeling method of directly predicting coordinates is not friendly to the prediction of neural network , The prediction accuracy is restricted to a certain extent .

.1、 First , In order to train better and more uniformly , The author hopes to present the picture with people at the core , Thus, a standardization method of joint coordinates is introduced .

So the normalized joint point coordinates ( Relative coordinates ) Shown by the following ：

The absolute position of the joint point coordinates predicted by the network relative to the picture is

  But I found that the rear attitude recognition Baseline It seems that this kind of standardization has not been adopted in the paper . One of my guesses is , at that time Batch Normalization(2015) Your article hasn't been published yet , therefore DeepPose(2014) According to the characteristics of attitude recognition, this standardized method is adopted . But actually , It and Batch Normalization The function may be similar , Therefore, in the subsequent network construction , If you do BN, Then there is no need to do such posture Standardization .

 2、 Basic network part

This one is not complicated , Reprinted as a whole AlexNet, Only the output is changed . From the original one-dimensional vector （ Long for classified “ Number of categories “） It becomes a one-dimensional vector （ Long for “ Double the number of joints “）. The reason why the output length is changed to twice as much Number of joints , Because each joint has Horizontal and vertical Two coordinates to predict .

It is worth noting that , Because the prediction coordinate is actually a highly nonlinear regression task , For convolutional networks, there is no good use of the spatial information of the image . So since 14 Year of Joint Training of a Convolutional Network and a Graphical Model for Human Pose Estimation After putting forward the heat map method , Mainstream networks rarely use the way of predicting coordinates for attitude recognition . The specific advantages and disadvantages of prediction coordinates and thermal diagram will also be introduced in subsequent articles .

The network structure in this paper ,Backbone Namely AlexNet

3、 Advanced network part

In order to achieve better Local observation effect , The author also establishes more for each joint in a small scale AlexNet To learn . Therefore, the implementation of this block is actually equivalent to extracting smaller picture frames for different joints , Then standardize these proposed local picture frames （ modular 1） and DNN（ modular 2） Re modeling and calculation of , The reason for this is that it is conducive to the improvement of accuracy in the process of enlarging the target position .

But compared with its implementation process , In fact, what is more important is the idea of this module , That's it As long as the joint position is extracted repeatedly in different stages , It is conducive to the improvement of accuracy . Such a combination local and global In fact, the idea of has been adopted in many subsequent papers （Stacked Hourglass/CPM/...）, And proved to be an effective idea .

  See ：

Human posture estimation -DeepPose ( Detailed instructions )

DeepPose summary - You know

DeepPose: Human posture estimation based on depth neural network Human Pose Estimation via Deep Neural Networkse

DeepPose Comprehensive analysis （ principle + Code +Colab） - You know

https://blog.csdn.net/header_zzj/article/details/116706719